Data processing method and apparatus, storage medium, and electronic device

Abstract

The application discloses a data processing method and device, a storage medium and an electronic device. The method comprises the following steps: acquiring a first forward output and a second forward output; processing the first forward output and the second forward output by a top model in a second participant to obtain a K-dimensional reasoning result; and in the case that the K is equal to L, performing dimension reduction on the K-dimensional reasoning result by a dimension reduction module in the second participant to obtain an M-dimensional reasoning result. The application solves the technical problem of low data security.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computers, and in particular, to a data processing method and device, a storage medium and an electronic device. BACKGROUND

[0002] In the related art, predicting user preferences can recommend media resources of interest to the user, such as video recommendations, product recommendations, and the like.

[0003] Based on user historical behavior data, a neural network model obtained through machine learning can predict user behavior. The larger the amount of training sample data, the higher the prediction accuracy of the neural network model. Therefore, the amount of training sample data has always been an important factor affecting the accuracy of user preference prediction. The user historical behavior data accumulated on a single platform is limited, and if multiple different platforms provide user historical behavior data together, the amount of training sample data can be increased, and thus the prediction accuracy of the neural network model can be improved. However, the multi-platform cooperation mode has the problem of data leakage, and it is difficult to guarantee the privacy and security of user data.

[0004] In view of the above problems, no effective solution has been proposed so far. SUMMARY

[0005] Embodiments of the present application provide a data processing method and device, a storage medium and an electronic device to at least solve the technical problem of low data security.

[0006] According to an aspect of some embodiments of the present application, a data processing method is provided, including: obtaining a first forward output and a second forward output, wherein the first forward output is an output obtained by processing first feature data by a first bottom model in a first participant, and the second forward output is an output obtained by processing second feature data by a second bottom model in a second participant; processing the first forward output and the second forward output by a top model in the second participant to obtain a K-dimensional inference result, wherein K is a positive integer greater than 2; in a case where K is equal to L, performing dimension reduction on the K-dimensional inference result by a dimension reduction module in the second participant to obtain an M-dimensional inference result, wherein M is a positive integer less than K and greater than or equal to 2, and L is a preset dimension threshold parameter; wherein the first bottom model, the second bottom model and the top model are obtained by jointly training a to-be-trained first bottom model in the first participant, a to-be-trained second bottom model in the second participant and a to-be-trained top model using a first sample feature data set, a second sample feature data set and a processed known sample label set, the processed known sample label set is a sample label set obtained by performing sample label processing on each M-dimensional known sample label in an M-dimensional known sample label set, and the sample label processing includes adding random noise.

[0007] Optionally, after the first forward output and the second forward output are obtained, the method further includes: fusing the first forward output and the second forward output by a split layer in the second participant to obtain a fusion result; and processing the first forward output and the second forward output by the top model in the second participant to obtain the K-dimensional inference result includes processing the fusion result by the top model in the second participant to obtain the K-dimensional inference result.

[0008] Optionally, the fusing the first forward output and the second forward output by the split layer in the second participant to obtain a fusion result includes: splicing the first forward output and the second forward output by the split layer in the second participant to obtain the fusion result; or performing mean value processing on the first forward output and the second forward output by the split layer in the second participant to obtain the fusion result.

[0009] Optionally, before the obtaining the first forward output and the second forward output, the method further comprises: obtaining a first sample feature data set, a second sample feature data set and an M-dimensional known sample label set having a corresponding relationship, wherein each M-dimensional known sample label in the M-dimensional known sample label set is used to represent an actual sample label; performing sample label processing on each M-dimensional known sample label in the M-dimensional known sample label set to obtain a processed known sample label set, wherein the sample label processing comprises: adding random noise; and performing joint training on a to-be-trained first bottom model in the first participant, a to-be-trained second bottom model in the second participant and a to-be-trained top model using the first sample feature data set, the second sample feature data set and the processed known sample label set until a first end condition is met, ending the training to obtain the first bottom model, the second bottom model and the top model.

[0010] Optionally, the performing sample label processing on each M-dimensional known sample label in the M-dimensional known sample label set to obtain a processed known sample label set comprises: in a case where the M is less than the L, performing dimension increasing on the M-dimensional known sample label set by a dimension increasing module to obtain a K-dimensional known sample label set; adding random noise to each K-dimensional known sample label in the K-dimensional known sample label set to obtain a K-dimensional known noise sample label set, wherein the processed known sample label set comprises the K-dimensional known noise sample label set; or in a case where the M is greater than or equal to the L, adding random noise to each M-dimensional known sample label in the M-dimensional known sample label set to obtain an M-dimensional known noise sample label set, wherein the processed known sample label set comprises the M-dimensional known noise sample label set.

[0011] Optionally, the adding random noise to each of the K-dimensional known sample label in the set of K-dimensional known sample labels to obtain a set of K-dimensional known noise sample labels comprises: adding random distribution noise to each of the K-dimensional known sample label in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels; or adding random opposite label noise to each of the K-dimensional known sample label in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels; or adding random opposite label and random distribution noise to each of the K-dimensional known sample label in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels; adding random noise to each of the M-dimensional known sample label in the set of M-dimensional known sample labels to obtain a set of M-dimensional known noise sample labels comprises: adding random distribution noise to each of the M-dimensional known sample label in the set of M-dimensional known sample labels to obtain the set of M-dimensional known noise sample labels; or adding random opposite label noise to each of the M-dimensional known sample label in the set of M-dimensional known sample labels to obtain the set of M-dimensional known noise sample labels; or adding random opposite label and random distribution noise to each of the M-dimensional known sample label in the set of M-dimensional known sample labels to obtain the set of M-dimensional known noise sample labels.

[0012] Optionally, the adding random distribution noise to each of the K-dimensional known sample label in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels comprises: performing the following steps on each of the K-dimensional known sample label in the set of K-dimensional known sample labels, wherein each of the K-dimensional known sample label in the set of K-dimensional known sample labels is a current K-dimensional known sample label: generating a K-dimensional random noise by a specified random distribution, processing the K-dimensional random noise by a softmax function to obtain a K-dimensional noise sample label, and determining a sum of the K-dimensional noise sample label and the current K-dimensional known sample label as a K-dimensional known noise sample label of the current K-dimensional known sample label. Optionally, the adding random distribution noise to each of the K-dimensional known sample label in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels comprises: performing the following steps on each of the K-dimensional known sample label in the set of K-dimensional known sample labels, wherein each of the K-dimensional known sample label in the set of K-dimensional known sample labels is a current K-dimensional known sample label: generating a K-dimensional random noise by a specified random distribution, processing the K-dimensional random noise by a softmax function to obtain a K-dimensional noise sample label, and determining a sum of the K-dimensional noise sample label and the current K-dimensional known sample label as a K-dimensional known noise sample label of the current K-dimensional known sample label. Optionally, the adding random distribution noise to each of the K-dimensional known sample label in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels comprises: performing the following steps on each of the K-dimensional known sample label in the set of K-dimensional known sample labels, wherein each of the K-dimensional known sample label in the set of K-dimensional known sample labels is a current K-dimensional known sample label: generating a K-dimensional random noise by a specified random distribution, processing the K-dimensional random noise by a softmax function to obtain a K-dimensional noise sample label, and determining a sum of the K-dimensional noise sample label and the current K-dimensional known sample label as a K-dimensional known noise sample label of the current K-dimensional known sample label. Optionally, the adding random distribution noise to each of the K-dimensional known sample label in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels comprises: performing the following steps on each of the K-dimensional known sample label in the set of K-dimensional known sample labels, wherein each of the K-dimensional known sample label in the set of K-dimensional known sample labels is a current K-dimensional known sample label: generating a K-dimensional random noise by a specified random distribution, processing the K-dimensional random noise by a softmax function to obtain a K-dimensional noise sample label, and determining a sum of the K-dimensional noise sample label and the current K-dimensional known sample label as a K-dimensional known noise sample label of the current K-dimensional known sample label. Optionally, the adding random distribution noise to each of the K-dimensional known sample label in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels comprises: performing the following steps on each of the K-dimensional known sample label in the set of K-dimensional known sample labels, wherein each of the K-dimensional known sample label in the set of K-dimensional known sample labels is a current K-dimensional known sample label: generating a K-dimensional random noise by a specified random distribution, processing the K-dimensional random noise by a softmax function to obtain a K-dimensional noise sample label, and determining a sum of the K-dimensional noise sample label and the current K-dimensional known sample label as a K-dimensional known noise sample label of the current K-dimensional known sample label.

[0013] Optionally, the K-dimensional known sample label set is added with random opposite label noise to obtain the K-dimensional known noise sample label set, including: performing the following steps on each K-dimensional known sample label in the K-dimensional known sample label set, and each K-dimensional known sample label in the performing of the following steps is a current K-dimensional known sample label: obtaining the opposite label of the current K-dimensional known sample label to obtain a K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; determining the K-dimensional opposite sample label randomly selected from the K-dimensional opposite sample label set as the K-dimensional known noise sample label of the current K-dimensional known sample label. Optionally, the K-dimensional known sample label set is added with random opposite label noise to obtain the K-dimensional known noise sample label set, including: performing the following steps on each K-dimensional known sample label in the K-dimensional known sample label set, and each K-dimensional known sample label in the performing of the following steps is a current K-dimensional known sample label: obtaining the opposite label of the current K-dimensional known sample label to obtain a K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; determining the K-dimensional opposite sample label randomly selected from the K-dimensional opposite sample label set as the K-dimensional known noise sample label of the current K-dimensional known sample label. Optionally, the K-dimensional known sample label set is added with random opposite label noise to obtain the K-dimensional known noise sample label set, including: performing the following steps on each K-dimensional known sample label in the K-dimensional known sample label set, and each K-dimensional known sample label in the performing of the following steps is a current K-dimensional known sample label: obtaining the opposite label of the current K-dimensional known sample label to obtain a K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; determining the K-dimensional opposite sample label randomly selected from the K-dimensional opposite sample label set as the K-dimensional known noise sample label of the current K-dimensional known sample label.

[0014] Optionally, the K-dimensional known sample label set is added with random opposite label noise to obtain the K-dimensional known noise sample label set, including: performing the following steps on each K-dimensional known sample label in the K-dimensional known sample label set, and each K-dimensional known sample label in the performing of the following steps is a current K-dimensional known sample label: obtaining the opposite label of the current K-dimensional known sample label to obtain a K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; determining the K-dimensional opposite sample label randomly selected from the K-dimensional opposite sample label set as the K-dimensional known noise sample label of the current K-dimensional known sample label. Optionally, the K-dimensional known sample label set is added with random opposite label noise to obtain the K-dimensional known noise sample label set, including: performing the following steps on each K-dimensional known sample label in the K-dimensional known sample label set, and each K-dimensional known sample label in the performing of the following steps is a current K-dimensional known sample label: obtaining the opposite label of the current K-dimensional known sample label to obtain a K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; determining the K-dimensional opposite sample label randomly selected from the K-dimensional opposite sample label set as the K-dimensional known noise sample label of the current K-dimensional known sample label. Optionally, the K-dimensional known sample label set is added with random opposite label noise to obtain the K-dimensional known noise sample label set, including: performing the following steps on each K-dimensional known sample label in the K-dimensional known sample label set, and each K-dimensional known sample label in the performing of the following steps is a current K-dimensional known sample label: obtaining the opposite label of the current K-dimensional known sample label to obtain a K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; determining the K-dimensional opposite sample label randomly selected from the K-dimensional opposite sample label set as the K-dimensional known noise sample label of the current K-dimensional known sample label.

[0015] Optionally, the K-dimensional known sample label set is added with random opposite label noise to obtain the K-dimensional known noise sample label set, including: performing the following steps on each K-dimensional known sample label in the K-dimensional known sample label set, and each K-dimensional known sample label in the performing of the following steps is a current K-dimensional known sample label: obtaining the opposite label of the current K-dimensional known sample label to obtain a K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; determining the K-dimensional opposite sample label randomly selected from the K-dimensional opposite sample label set as the K-dimensional known noise sample label of the current K-dimensional known sample label.

[0016] In a case that the processed known sample label set is the K-dimensional known noise sample label set, the first to-be-trained bottom model, the second to-be-trained bottom model and the to-be-trained top model are trained for the jth round, where j is a positive integer greater than or equal to 1, and the first bottom model, the second bottom model and the top model obtained by the 0th round of training are the to-be-trained first bottom model, the to-be-trained second bottom model and the to-be-trained top model, and the training comprises the following steps:

[0017] S first sample feature data used in the jth round of the first sample feature data set are input into the first bottom model trained in the (j-1)th round to obtain S first sample forward outputs output by the jth round of training, where S is a positive integer greater than or equal to 1; S second sample feature data used in the jth round of the second sample feature data set are input into the second bottom model trained in the (j-1)th round to obtain S second sample forward outputs output by the jth round of training; the S first sample forward outputs output by the jth round of training and the S second sample forward outputs output by the jth round of training are input into the top model trained in the (j-1)th round to obtain S K-dimensional sample inference results output by the jth round of training; in a case that the first end condition is met between the S K-dimensional sample inference results output by the jth round of training and the S K-dimensional known noise sample labels used in the jth round, the training is ended to obtain the first bottom model, the second bottom model and the top model, and the K-dimensional known noise sample labels used in the jth round have the corresponding relationship with the first sample feature data used in the jth round and the second sample feature data used in the jth round.

[0018] Optionally, the method further comprises: in a case that the first end condition is not met between the S K-dimensional sample inference results output by the jth round of training and the S K-dimensional known noise sample labels used in the jth round, adjusting model parameters of the first bottom model trained in the (j-1)th round, the second bottom model and the top model trained in the (j-1)th round by back propagation.

[0019] Optionally, the adjusting of the model parameters of the top model trained in the (j-1)th round by back propagation comprises: the second participant obtains S first gradients from S loss functions between the S K-dimensional sample inference results and the S K-dimensional known noise sample labels and the S K-dimensional sample inference results, and adjusts the model parameters of the top model trained in the (j-1)th round by the S first gradients.

[0020] The adjusting the model parameters of the second base model obtained by the j-1th round of training through back propagation comprises: the second participant obtaining S second gradients through S loss functions between the S K-dimensional sample inference results and the S K-dimensional known noise sample labels and the model parameters of the second base model obtained by the j-1th round of training, and adjusting the model parameters of the second base model obtained by the j-1th round of training through the S second gradients.

[0021] The adjusting the first base model obtained by the j-1th round of training through back propagation comprises: the second participant obtaining S third gradients through S loss functions between the S K-dimensional sample inference results and the S K-dimensional known noise sample labels and the S first sample forward outputs; and the second participant sending the S third gradients to the first participant, and the first participant adjusting the first base model obtained by the j-1th round of training through the S third gradients.

[0022] Optionally, the dimension increasing module is configured to increase the dimension of the M-dimensional known sample label set, comprising: obtaining K K-dimensional one-hot encoding vectors, randomly dividing the K K-dimensional one-hot encoding vectors into M groups, or dividing the K K-dimensional one-hot encoding vectors into M groups according to a preset rule, to obtain M groups of K-dimensional one-hot encoding vectors; and mapping the M-dimensional known sample labels to the M groups of K-dimensional one-hot encoding vectors respectively to obtain K-dimensional known sample labels corresponding to the M-dimensional known sample labels.

[0023] The dimension decreasing module in the second participant is configured to decrease the dimension of the K-dimensional inference result, comprising: performing inner product operation on the K-dimensional inference result and M K-dimensional decoding vectors to obtain the M-dimensional inference result, wherein the M K-dimensional decoding vectors are decoding vectors obtained according to the M groups of K-dimensional one-hot encoding.

[0024] According to another aspect of the embodiments of the present application, a data processing method is further provided, comprising: obtaining a first sample feature data set, a second sample feature data set and an M-dimensional known sample label set having a corresponding relationship, wherein each M-dimensional known sample label in the M-dimensional known sample label set is used to represent an actual sample label; performing sample label processing on each M-dimensional known sample label in the M-dimensional known sample label set to obtain a processed known sample label set, wherein the sample label processing comprises: adding random noise; and using the first sample feature data set, the second sample feature data set and the processed known sample label set to jointly train a first base model to be trained in a first participant, a second base model to be trained in a second participant and a top model to be trained until a first end condition is met, ending the training to obtain a first base model, a second base model and a top model.

[0025] Optionally, the sample label processing on each of the M-dimensional known sample labels in the set of M-dimensional known sample labels comprises, in the case that the M is less than the L, dimension increasing on the set of M-dimensional known sample labels by a dimension increasing module to obtain a set of K-dimensional known sample labels; adding random noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain a set of K-dimensional known noise sample labels, wherein the set of processed known sample labels comprises the set of K-dimensional known noise sample labels; or, in the case that the M is greater than or equal to the L, adding random noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain a set of M-dimensional known noise sample labels, wherein the set of processed known sample labels comprises the set of M-dimensional known noise sample labels.

[0026] Optionally, the adding random noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain a set of K-dimensional known noise sample labels comprises: adding randomly distributed noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels; or adding randomly opposite label noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels; or adding randomly opposite label and randomly distributed noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels; the adding random noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain a set of M-dimensional known noise sample labels comprises: adding randomly distributed noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain the set of M-dimensional known noise sample labels; or adding randomly opposite label noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain the set of M-dimensional known noise sample labels; or adding randomly opposite label and randomly distributed noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain the set of M-dimensional known noise sample labels.

[0027] Optionally, the adding randomly distributed noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels comprises: performing the following steps on each of the K-dimensional known sample labels in the set of K-dimensional known sample labels, wherein each of the K-dimensional known sample labels in the performing of the following steps is a current K-dimensional known sample label: randomly generating random noise of the K dimension by a specified random distribution, and adding the random noise of the K dimension to the current K-dimensional known sample label by a softmax function to obtain a current K-dimensional known noise sample label; and adding the current K-dimensional known noise sample label to the set of K-dimensional known noise sample labels; or the adding randomly distributed noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain the set of M-dimensional known noise sample labels comprises: performing the following steps on each of the M-dimensional known sample labels in the set of M-dimensional known sample labels, wherein each of the M-dimensional known sample labels in the performing of the following steps is a current M-dimensional known sample label: randomly generating random noise of the M dimension by a specified random distribution, and adding the random noise of the M dimension to the current M-dimensional known sample label by a softmax function to obtain a current M-dimensional known noise sample label; and adding the current M-dimensional known noise sample label to the set of M-dimensional known noise sample labels. ​ a K-dimensional random noise of the current K-dimensional known sample label to obtain a K-dimensional known noise sample label of the current K-dimensional known sample label. a K-dimensional random noise of the current K-dimensional known sample label to obtain a K-dimensional known noise sample label of the current K-dimensional known sample label. a K-dimensional random noise of the current K-dimensional known sample label to obtain a K-dimensional known noise sample label of the current K-dimensional known sample label.

[0028] Optionally, the K-dimensional known noise sample label set is obtained by adding a random noise of an opposite label to each K-dimensional known sample label in the K-dimensional known sample label set, including: performing the following steps on each K-dimensional known sample label in the K-dimensional known sample label set, and each K-dimensional known sample label in the performing is a current K-dimensional known sample label: obtaining an opposite label of the current K-dimensional known sample label to obtain a K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; and determining a K-dimensional opposite sample label randomly selected from the K-dimensional opposite sample label set as a a K-dimensional random noise of the current K-dimensional known sample label to obtain a K-dimensional known noise sample label of the current K-dimensional known sample label. a K-dimensional random noise of the current K-dimensional known sample label to obtain a K-dimensional known noise sample label of the current K-dimensional known sample label.

[0029] Optionally, the K-dimensional known noise sample label set is obtained by adding a random noise of an opposite label and a random distribution to each K-dimensional known sample label in the K-dimensional known sample label set, including: performing the following steps on each K-dimensional known sample label in the K-dimensional known sample label set, and each K-dimensional known sample label in the performing is a current K-dimensional known sample label: randomly generating a random disturbance of the current K-dimensional known sample label through a specified random distribution; obtaining an opposite label of the current K-dimensional known sample label to obtain a K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; and determining a product of a K-dimensional opposite sample label randomly selected from the K-dimensional opposite sample label set and the random disturbance of the current K-dimensional known sample label as a a K-dimensional random noise of the current K-dimensional known sample label to obtain a K-dimensional known noise sample label of the current K-dimensional known sample label. a K-dimensional random noise of the current K-dimensional known sample label to obtain a K-dimensional known noise sample label of the current K-dimensional known sample label.

[0030] Optionally, the joint training of the first bottom model to be trained in the first participant, the second bottom model to be trained in the second participant, and the top model to be trained includes:

[0031] In a case that the processed known sample label set is the K-dimensional known noise sample label set, the first to-be-trained bottom model, the second to-be-trained bottom model and the to-be-trained top model are trained for the jth round, where j is a positive integer greater than or equal to 1, and the first bottom model, the second bottom model and the top model obtained by the 0th round of training are the to-be-trained first bottom model, the to-be-trained second bottom model and the to-be-trained top model, and the training comprises the following steps:

[0032] S first sample feature data used in the jth round of the first sample feature data set are input into the first bottom model trained in the (j-1)th round to obtain S first sample forward outputs output by the jth round of training, where S is a positive integer greater than or equal to 1; S second sample feature data used in the jth round of the second sample feature data set are input into the second bottom model trained in the (j-1)th round to obtain S second sample forward outputs output by the jth round of training; the S first sample forward outputs output by the jth round of training and the S second sample forward outputs output by the jth round of training are input into the top model trained in the (j-1)th round to obtain S K-dimensional sample inference results output by the jth round of training; in a case that the first end condition is met between the S K-dimensional sample inference results output by the jth round of training and the S K-dimensional known noise sample labels used in the jth round, the training is ended to obtain the first bottom model, the second bottom model and the top model, and the K-dimensional known noise sample labels used in the jth round have the corresponding relationship with the first sample feature data used in the jth round and the second sample feature data used in the jth round.

[0033] Optionally, the method further comprises: in a case that the first end condition is not met between the S K-dimensional sample inference results output by the jth round of training and the S K-dimensional known noise sample labels used in the jth round, adjusting model parameters of the first bottom model trained in the (j-1)th round, the second bottom model and the top model trained in the (j-1)th round by back propagation.

[0034] Optionally, the adjusting the model parameters of the top model trained in the (j-1)th round by back propagation comprises: the second participant obtains S first gradients from S loss functions between the S K-dimensional sample inference results and the S K-dimensional known noise sample labels and the S K-dimensional sample inference results, and adjusts the model parameters of the top model trained in the (j-1)th round by the S first gradients.

[0035] The adjusting the model parameters of the second base model obtained through the j-1th round of training by back propagation comprises: the second participant obtaining S second gradients through S loss functions between the S K-dimensional sample inference results and the S K-dimensional known noise sample labels and the model parameters of the second base model obtained through the j-1th round of training, and adjusting the model parameters of the second base model obtained through the j-1th round of training through the S second gradients.

[0036] The adjusting the first base model obtained through the j-1th round of training by back propagation comprises: the second participant obtaining S third gradients through S loss functions between the S K-dimensional sample inference results and the S K-dimensional known noise sample labels and the S first sample forward outputs; and the second participant sending the S third gradients to the first participant, and the first participant adjusting the first base model obtained through the j-1th round of training through the S third gradients.

[0037] Optionally, the dimension increasing module is configured to increase the dimension of the M-dimensional known sample label set, comprising: obtaining K K-dimensional one-hot encoding vectors, randomly dividing the K K-dimensional one-hot encoding vectors into M groups, or dividing the K K-dimensional one-hot encoding vectors into M groups according to a preset rule, to obtain M groups of K-dimensional one-hot encoding vectors; and mapping the M-dimensional known sample labels to the M groups of K-dimensional one-hot encoding vectors respectively, to obtain K-dimensional known sample labels corresponding to the M-dimensional known sample labels.

[0038] According to another aspect of the embodiments of the present application, a data processing method is further provided, comprising: obtaining a first forward output, wherein the first forward output is an output obtained by processing first feature data by a first base model in a first participant; processing the first forward output by a top model in a second participant to obtain a K-dimensional inference result, wherein the K is a positive integer greater than 2; in a case where the K is equal to L, processing the K-dimensional inference result by a dimension reduction module in the second participant to obtain an M-dimensional inference result, wherein the M is a positive integer less than the K and greater than or equal to 2, and the L is a preset dimension threshold parameter; wherein the first base model and the top model are obtained by jointly training a to-be-trained first base model in the first participant and a to-be-trained top model in the second participant using a first sample feature data set and a processed known sample label set, the processed known sample label set is a sample label set obtained by processing each M-dimensional known sample label in an M-dimensional known sample label set, and the sample label processing comprises adding random noise.

[0039] Optionally, before the acquiring the first forward output, the method further comprises: acquiring a first sample feature data set having a corresponding relationship, and an M-dimensional known sample label set, wherein each M-dimensional known sample label in the M-dimensional known sample label set is used to represent an actual sample label; performing sample label processing on each M-dimensional known sample label in the M-dimensional known sample label set to obtain a processed known sample label set, wherein the sample label processing comprises: adding random noise; and using the first sample feature data set and the processed known sample label set to jointly train a to-be-trained first bottom model in the first participant and a to-be-trained top model in the second participant until a first end condition is met, ending the training, and obtaining the first bottom model and the top model.

[0040] Optionally, the performing sample label processing on each M-dimensional known sample label in the M-dimensional known sample label set to obtain a processed known sample label set comprises: in a case where the M is less than the L, performing dimensionality increasing on the M-dimensional known sample label set by a dimensionality increasing module to obtain a K-dimensional known sample label set; adding random noise to each K-dimensional known sample label in the K-dimensional known sample label set to obtain a K-dimensional known noise sample label set, wherein the processed known sample label set comprises the K-dimensional known noise sample label set; or in a case where the M is greater than or equal to the L, adding random noise to each M-dimensional known sample label in the M-dimensional known sample label set to obtain an M-dimensional known noise sample label set, wherein the processed known sample label set comprises the M-dimensional known noise sample label set.

[0041] Optionally, the adding random noise to each K-dimensional known sample label in the K-dimensional known sample label set to obtain a K-dimensional known noise sample label set comprises: adding randomly distributed noise to each K-dimensional known sample label in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set; or adding randomly opposite label noise to each K-dimensional known sample label in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set; or adding randomly opposite label and randomly distributed noise to each K-dimensional known sample label in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set.

[0042] Adding random noise to each of the M-dimensional known sample labels in the M-dimensional known sample label set to obtain an M-dimensional known noise sample label set includes: adding randomly distributed noise to each of the M-dimensional known sample labels in the M-dimensional known sample label set to obtain the M-dimensional known noise sample label set; or, adding noise of randomly opposite labels to each of the M-dimensional known sample labels in the M-dimensional known sample label set to obtain the M-dimensional known noise sample label set; or, adding random opposite labels and randomly distributed noise to each of the M-dimensional known sample labels in the M-dimensional known sample label set to obtain the M-dimensional known noise sample label set.

[0043] Optionally, adding randomly distributed noise to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set includes: performing the following steps on each of the K-dimensional known sample labels in the K-dimensional known sample label set, wherein each of the K-dimensional known sample labels during the execution of the following steps is the current K-dimensional known sample label: randomly generated through a specified random distribution. Dimensional random noise, and applied to the softmax function. The random noise of dimension is processed to obtain 3D random noise; The sum of the random noise and the current K-dimensional known sample label is determined as the K-dimensional known noise sample label of the current K-dimensional known sample label.

[0044] Optionally, adding random opposing labels to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set includes: performing the following steps on each of the K-dimensional known sample labels in the K-dimensional known sample label set, where each of the K-dimensional known sample labels during the following steps is the current K-dimensional known sample label: obtaining the opposing label of the current K-dimensional known sample label to obtain a K-dimensional opposing sample label set corresponding to the current K-dimensional known sample label; and determining the randomly selected K-dimensional opposing sample label from the K-dimensional opposing sample label set as the current K-dimensional known sample label. 3D random noise; The sum of the random noise and the current K-dimensional known sample label is determined as the K-dimensional known noise sample label of the current K-dimensional known sample label.

[0045] Optionally, adding random opposing labels and randomly distributed noise to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set includes: performing the following steps on each of the K-dimensional known sample labels in the K-dimensional known sample label set, where each of the K-dimensional known sample labels during the following steps is the current K-dimensional known sample label: randomly generating a random perturbation of the current K-dimensional known sample label through a specified random distribution; obtaining the opposing label of the current K-dimensional known sample label to obtain a K-dimensional opposing sample label set corresponding to the current K-dimensional known sample label; and determining the product of the randomly selected K-dimensional opposing sample label from the K-dimensional opposing sample label set and the random perturbation of the current K-dimensional known sample label as the current K-dimensional known sample label. 3D random noise; The sum of the random noise and the current K-dimensional known sample label is determined as the K-dimensional known noise sample label of the current K-dimensional known sample label.

[0046] Optionally, the M-dimensional known sample label set is upgraded using a dimensionality-upgrading module, including: obtaining K K-dimensional one-hot encoded vectors; randomly dividing the K K-dimensional one-hot encoded vectors into M groups, or dividing them into M groups according to a preset rule, to obtain M groups of K-dimensional one-hot encoded vectors; and mapping the M-dimensional known sample labels to the M groups of K-dimensional one-hot encoded vectors to obtain K-dimensional known sample labels corresponding to the M-dimensional known sample labels.

[0047] According to another aspect of the embodiments of this application, a data processing method is also provided, comprising: acquiring a first sample feature data set with corresponding relationships and an M-dimensional known sample label set, wherein each M-dimensional known sample label in the M-dimensional known sample label set is used to represent an actual sample label; performing sample label processing on each M-dimensional known sample label in the M-dimensional known sample label set to obtain a processed known sample label set, wherein the sample label processing includes: adding random noise; using the first sample feature data set and the processed known sample label set to jointly train a first bottom model to be trained in the first participant and a top model to be trained in the second participant until a first termination condition is met, thereby ending the training and obtaining the first bottom model and the top model.

[0048] Optionally, the sample label processing on each of the M-dimensional known sample labels in the set of M-dimensional known sample labels comprises: in the case that the M is less than the L, dimension increasing on the set of M-dimensional known sample labels by a dimension increasing module to obtain a set of K-dimensional known sample labels; adding random noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain a set of K-dimensional known noise sample labels, wherein the set of processed known sample labels comprises the set of K-dimensional known noise sample labels; or in the case that the M is greater than or equal to the L, adding random noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain a set of M-dimensional known noise sample labels, wherein the set of processed known sample labels comprises the set of M-dimensional known noise sample labels.

[0049] Optionally, the adding random noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain a set of K-dimensional known noise sample labels comprises: adding randomly distributed noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels; or adding randomly opposite label noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels; or adding randomly opposite label and randomly distributed noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels.

[0050] The adding random noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain a set of M-dimensional known noise sample labels comprises: adding randomly distributed noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain the set of M-dimensional known noise sample labels; or adding randomly opposite label noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain the set of M-dimensional known noise sample labels; or adding randomly opposite label and randomly distributed noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain the set of M-dimensional known noise sample labels.

[0051] Optionally, the adding randomly distributed noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels comprises: performing the following steps on each of the K-dimensional known sample labels in the set of K-dimensional known sample labels, wherein each of the K-dimensional known sample labels in the performing of the following steps is a current K-dimensional known sample label: randomly generating a K-dimensional noise sample label by a specified random distribution; and adding the K-dimensional noise sample label to the current K-dimensional known sample label to obtain a K-dimensional noise sample label. a K-dimensional random noise, and processing the K-dimensional random noise by a softmax function to obtain a K-dimensional noise sample label of the current K-dimensional known sample label. a K-dimensional random noise, and processing the K-dimensional random noise by a softmax function to obtain a K-dimensional noise sample label of the current K-dimensional known sample label. a K-dimensional random noise, and processing the K-dimensional random noise by a softmax function to obtain a K-dimensional noise sample label of the current K-dimensional known sample label. a K-dimensional random noise, and processing the K-dimensional random noise by a softmax function to obtain a K-dimensional noise sample label of the current K-dimensional known sample label.

[0052] Optionally, a random noise of an opposite label is added to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set, including: performing the following steps on each of the K-dimensional known sample labels in the K-dimensional known sample label set, and each of the K-dimensional known sample labels in the performing of the following steps is a current K-dimensional known sample label: obtaining an opposite label of the current K-dimensional known sample label to obtain a K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; and determining a K-dimensional opposite sample label randomly selected from the K-dimensional opposite sample label set as a K-dimensional known noise sample label of the current K-dimensional known sample label. a K-dimensional random noise, and processing the K-dimensional random noise by a softmax function to obtain a K-dimensional noise sample label of the current K-dimensional known sample label. a K-dimensional random noise, and processing the K-dimensional random noise by a softmax function to obtain a K-dimensional noise sample label of the current K-dimensional known sample label.

[0053] Optionally, a random noise of an opposite label and a random distribution are added to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set, including: performing the following steps on each of the K-dimensional known sample labels in the K-dimensional known sample label set, and each of the K-dimensional known sample labels in the performing of the following steps is a current K-dimensional known sample label: randomly generating a random disturbance of the current K-dimensional known sample label by a specified random distribution; obtaining an opposite label of the current K-dimensional known sample label to obtain a K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; and determining a product of a K-dimensional opposite sample label randomly selected from the K-dimensional opposite sample label set and the random disturbance of the current K-dimensional known sample label as a K-dimensional known noise sample label of the current K-dimensional known sample label. a K-dimensional random noise, and processing the K-dimensional random noise by a softmax function to obtain a K-dimensional noise sample label of the current K-dimensional known sample label. a K-dimensional random noise, and processing the K-dimensional random noise by a softmax function to obtain a K-dimensional noise sample label of the current K-dimensional known sample label.

[0054] Optionally, the M-dimensional known sample label set is dimensionally increased by a dimension increasing module, including: obtaining K K-dimensional one-hot encoding vectors, randomly dividing the K K-dimensional one-hot encoding vectors into M groups, or dividing the K K-dimensional one-hot encoding vectors into M groups according to a preset rule, to obtain M groups of K-dimensional one-hot encoding vectors; and mapping the M-dimensional known sample labels to the M groups of K-dimensional one-hot encoding vectors respectively, to obtain K-dimensional known sample labels corresponding to the M-dimensional known sample labels.

[0055] According to another aspect of the embodiments of the present application, a data processing apparatus is further provided, including: a first obtaining module, configured to obtain a first forward output and a second forward output, wherein the first forward output is an output obtained by processing first feature data by a first bottom model in a first participant, and the second forward output is an output obtained by processing second feature data by a second bottom model in a second participant; a first processing module, configured to process the first forward output and the second forward output by a top model in the second participant, to obtain a K-dimensional inference result, wherein K is a positive integer greater than 2; and a first dimension reducing module, configured to, in a case where the K is equal to L, reduce the dimension of the K-dimensional inference result by a dimension reducing module in the second participant, to obtain an M-dimensional inference result, wherein M is a positive integer less than K and greater than or equal to 2, and L is a preset dimension threshold parameter; wherein the first bottom model, the second bottom model and the top model are obtained by jointly training a to-be-trained first bottom model in the first participant, a to-be-trained second bottom model in the second participant and a to-be-trained top model, using a first sample feature data set, a second sample feature data set and a processed known sample label set, the processed known sample label set is a sample label set obtained by performing sample label processing on each M-dimensional known sample label in an M-dimensional known sample label set, and the sample label processing includes adding random noise.

[0056] According to another aspect of the embodiments of the present application, a data processing apparatus is further provided, comprising: a second obtaining module, configured to obtain a first sample feature data set, a second sample feature data set and an M-dimensional known sample label set having a corresponding relationship, wherein each M-dimensional known sample label in the M-dimensional known sample label set is used to represent an actual sample label; a second processing module, configured to perform sample label processing on each M-dimensional known sample label in the M-dimensional known sample label set to obtain a processed known sample label set, wherein the sample label processing comprises adding random noise; and a first training module, configured to perform joint training on a first bottom model to be trained in a first participant, a second bottom model to be trained in a second participant and a top model to be trained using the first sample feature data set, the second sample feature data set and the processed known sample label set until a first end condition is met, and then end the training to obtain the first bottom model, the second bottom model and the top model.

[0057] According to another aspect of the embodiments of the present application, a data processing apparatus is further provided, comprising: a third obtaining module, configured to obtain a first forward output, wherein the first forward output is an output obtained by processing first feature data by a first bottom model in a first participant; a third processing module, configured to process the first forward output by a top model in the second participant to obtain a K-dimensional inference result, wherein K is a positive integer greater than 2; and a second dimension reduction module, configured to perform dimension reduction on the K-dimensional inference result by a dimension reduction module in the second participant to obtain an M-dimensional inference result in a case that the K is equal to L, wherein M is a positive integer less than the K and greater than or equal to 2, and L is a preset dimension threshold parameter; wherein the first bottom model and the top model are obtained by performing joint training on a first bottom model to be trained in the first participant and a top model to be trained in the second participant using a first sample feature data set and a processed known sample label set, the processed known sample label set is a sample label set obtained by performing sample label processing on each M-dimensional known sample label in an M-dimensional known sample label set, and the sample label processing comprises adding random noise.

[0058] According to another aspect of the embodiments of the present application, a data processing apparatus is further provided, comprising: a fourth acquisition module configured to acquire a first sample feature data set having a corresponding relationship, and an M-dimensional known sample label set, wherein each M-dimensional known sample label in the M-dimensional known sample label set is used to represent an actual sample label; a fourth processing module configured to perform sample label processing on each M-dimensional known sample label in the M-dimensional known sample label set to obtain a processed known sample label set, wherein the sample label processing comprises adding random noise; and a second training module configured to perform joint training on a first bottom model to be trained in the first participant and a top model to be trained in the second participant using the first sample feature data set and the processed known sample label set until a first end condition is met, and then ending the training to obtain the first bottom model and the top model.

[0059] According to still another aspect of the embodiments of the present application, a computer readable storage medium is further provided, and the computer readable storage medium stores a computer program. The computer program is configured to execute the above data processing method when running.

[0060] According to still another aspect of the embodiments of the present application, a computer program product or computer program is provided, and the computer program product or computer program comprises computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions to enable the computer device to perform the above data processing method.

[0061] According to still another aspect of the embodiments of the present application, an electronic device is further provided, and the electronic device comprises a memory and a processor. The memory stores a computer program, and the processor is configured to execute the above data processing method through the computer program.

[0062] In the embodiments of the present application, the first participant provides first feature data, the second participant provides second feature data, the first bottom model is located in the first participant, and the second bottom model is located in the second participant. The first feature data is processed by the first bottom model to obtain a first forward output, the second feature data is processed by the second bottom model to obtain a second forward output, and the first forward output and the second forward output are processed by the top model in the second participant to obtain a K-dimensional inference result. In the case where K is equal to L, the K-dimensional inference result is dimensionally reduced by the dimension reduction module in the second participant to obtain an M-dimensional inference result. Since the M-dimensional inference result needs to be dimensionally reduced, the technical effect of improving data security is achieved, and the technical problem of low data security is solved.

[0063] In addition, in the process of jointly training the first bottom model to be trained in the first participant, the second bottom model to be trained in the second participant, and the top model to be trained, if M of the M-dimensional known sample label is less than the preset dimension threshold parameter L, the M-dimensional known sample label is added with random noise after dimension increasing, and if M of the M-dimensional known sample label is greater than or equal to the preset dimension threshold parameter L, the M-dimensional known sample label is added with random noise. In this way, since the second participant adds random noise to the M-dimensional known sample label provided by the second participant, the M-dimensional known sample label provided by the second participant is prevented from being leaked, and the technical effect of improving data security is achieved. BRIEF DESCRIPTION OF DRAWINGS

[0064] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and serve to explain the principles of the application, and do not limit the application in any way. In the drawings:

[0065] Figure 1 is a schematic diagram of an application environment of an optional data processing method according to an embodiment of the application;

[0066] Figure 2 is a flowchart of an optional data processing method according to an embodiment of the application;

[0067] Figure 3 is a schematic diagram of an optional longitudinal federated deep learning method according to an embodiment of the application;

[0068] Figure 4 is another optional inference stage schematic diagram according to an embodiment of the application;

[0069] Figure 5 is another optional longitudinal federated deep learning model training flowchart according to an embodiment of the application;

[0070] Figure 6 is another optional training stage schematic diagram according to an embodiment of the application;

[0071] Figure 7 is another optional federated deep model training flowchart according to an embodiment of the application;

[0072] Figure 8 is another optional federated deep model usage flowchart according to an embodiment of the application;

[0073] Figure 9 is a structural schematic diagram of an optional data processing apparatus according to an embodiment of the application;

[0074] Figure 10is a structural schematic diagram of another optional data processing apparatus according to an embodiment of the application;

[0075] Figure 11 is a computer system structural block diagram of an optional electronic device according to an embodiment of the application;

[0076] Figure 12 is a structural schematic diagram of an optional electronic device according to an embodiment of the application. DETAILED DESCRIPTION

[0077] In order to enable persons skilled in the art to better understand the scheme of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by persons skilled in the art without creative labor should be within the scope of protection of the present application.

[0078] It should be noted that the terms "first", "second" and the like in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0079] First, some of the nouns or terms appearing in the description of the embodiments of the present application are applicable to the following explanations:

[0080] Federated learning (FL for short);

[0081] Vertical federated learning (VFL for short);

[0082] Split learning (SL for short);

[0083] Label inference attack (LIA for short);

[0084] Clustering attack (CA for short);

[0085] Model Completion Attack (MCA).

[0086] The present application will be described in conjunction with embodiments below:

[0087] According to an aspect of the embodiments of the present application, a data processing method is provided. Optionally, as an optional implementation, the above-mentioned data processing method can be applied in an application environment as shown in Figure 1 , but is not limited to, the application environment. The application environment can include: a user device 101, a server 102, a server 103, and a user device 104, wherein the user device 101 and the server 102 are located at a second participant, and the user device 104 and the server 103 are located at a first participant. The first participant can also be referred to as a Host party, and the second participant can be referred to as a Guest party.

[0088] Optionally, in the present embodiment, the above-mentioned user device 101 and user device 104 can include, but are not limited to, at least one of the following: a mobile phone (such as an Android mobile phone, an iOS mobile phone, etc.), a notebook computer, a tablet computer, a palm computer, a MID (Mobile Internet Device), a PAD, a desktop computer, a smart television, a medical device, etc. The user device can be configured with a target client, which can be a game client, an instant messaging client, a browser client, a video client, a shopping client, etc.

[0089] Optionally, the above-mentioned network can include, but is not limited to: a wired network and a wireless network, wherein the wired network includes: a local area network, a metropolitan area network, and a wide area network, and the wireless network includes: Bluetooth, WIFI, and other wireless communication networks.

[0090] Optionally, the above-mentioned server 102 and server 103 can be a single server, or a server cluster composed of multiple servers, or a cloud server. The server 102 and the server 103 can be, but are not limited to: a database and a processing engine. The database can be used to store data, for example, the database in the server 102 can be used to store the first feature data, and the database in the server 103 can be used to store the second feature data. The processing engine processes data, for example, the processing engine in the server 102 can be used to process the first feature data through the first bottom model to obtain a first forward output, the processing engine in the server 103 can be used to process the second feature data through the second bottom model to obtain a second forward output, process the first forward output and the second forward output through the top model to obtain a K-dimensional inference result, and process the K-dimensional inference result through the dimension reduction module to obtain an M-dimensional inference result.

[0091] Optionally, in the embodiment, the above data processing method can also be implemented by a server, for example, Figure 1 implemented in the server 103 of the second participant as shown; or jointly implemented by the user equipment 104 and the server 103 of the second participant.

[0092] The above is only an example, and the embodiment is not specifically limited.

[0093] It can be understood that in the specific embodiments of the present application, the data related to user information (for example, first feature data, second feature data) is involved, and when the above embodiments of the present application are applied to specific products or technologies, the user's permission or consent is required, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards of relevant countries and regions.

[0094] Optionally, as an optional implementation, as shown in Figure 2 The above data processing method comprises:

[0095] Step S202, obtaining a first forward output and a second forward output, wherein the first forward output is an output obtained by processing first feature data by a first bottom model in a first participant, and the second forward output is an output obtained by processing second feature data by a second bottom model in a second participant;

[0096] The first feature data can be feature data of a target object, and the target object can be a target user. The first feature data can be portrait features (for example, gender, age, education, location, etc.) of the target user. The second feature data can be feature data of a target media resource, and the target media resource includes but is not limited to media resources such as video, audio, and goods. Taking a video as an example, the second feature data can be data such as the playing time of the video, the content of the video, and the creation location of the video.

[0097] Step S204, processing the first forward output and the second forward output by a top model in the second participant to obtain a K-dimensional inference result, wherein K is a positive integer greater than 2;

[0098] Step S206, in the case where K is equal to L, dimension reduction is performed on the K-dimensional inference result by a dimension reduction module in the second participant to obtain an M-dimensional inference result, wherein M is a positive integer less than K and greater than or equal to 2, and L is a preset dimension threshold parameter;

[0099] The first bottom model, the second bottom model and the top model are obtained by jointly training a to-be-trained first bottom model in the first participant, a to-be-trained second bottom model in the second participant and a to-be-trained top model using a first sample feature data set, a second sample feature data set and a processed known sample label set, the processed known sample label set is a sample label set obtained by performing sample label processing on each M-dimensional known sample label in an M-dimensional known sample label set, and the sample label processing includes adding random noise

[0100] The M-dimensional inference result is used to represent a predicted operation result, and the operation result is used to represent whether the target object performs a target operation on the target media resource, for example, whether a target user clicks to play a video, or whether a target user will purchase a product.

[0101] As an optional implementation, federated learning (FL) technology trains an effective machine learning model through interaction of models or intermediate results, can avoid direct transmission of data, and provides a new solution for cross-department, cross-platform, cross-organization and cross-industry data cooperation. According to the distribution of data among different participants, federated learning can be divided into vertical federated learning (VFL), horizontal federated learning (HFL) and federated transfer learning (FTL). Split learning (SL) is a new vertical federated learning paradigm, which realizes federated deep learning model training without data exchange through message interaction of a split layer (also called a cut layer or an interactive layer). The present application can be applied to vertical federated learning (VFL) and split learning (SL) scenarios.

[0102] Vertical federated learning increases the amount of features participating in training by jointly using multiple participants who have the same sample space and different feature space data sets, thereby improving the training effect of the model. In this scenario, the Host party (i.e., the first participant) and the Guest party (i.e., the second participant) have the same sample space. To train an inference model, the Guest party wants to cooperate with the feature data of the Host party without exposing private information, while ensuring that its own label is not leaked.

[0103] As Figure 3As shown, the core idea of vertical federated deep learning is to let each participant train a bottom model using its own features. During the training process, the guest party inputs the output features of all bottom models into the top model after fusion (can be spliced or calculate the average, etc.), and uses the data label and the output of the top model to calculate the loss function and back-propagate the gradient to update the parameters of the bottom model. Split learning can be regarded as a special form of vertical federated deep learning. Whether it is vertical federated learning or split learning, the data of the host party and the guest party is not out of the domain and not shared, and the data is available but invisible, thereby strengthening the protection of data privacy.

[0104] However, during the training process, the host party can implement a label inference attack (LIA) against the guest party by analyzing the gradient of the split layer. Because this attack is launched by the host party, it is independent of the existence of the data features and the bottom model of the guest party, and therefore both vertical federated learning and split learning can be threatened by label inference attacks. In the label protection scheme implemented by the guest party, isotropic Gaussian noise is added to the gradient of the back propagation, or optimized perturbation noise is added to the gradient of the back propagation, or an autoencoder (AE) is trained to encode the label information of the guest party, or a dimension conversion mechanism is used to train the classification task in a higher dimension.

[0105] The scheme based on isotropic Gaussian noise causes the model to perform worse on the main task because the noise and the original model gradient are in different distributions. The defense mechanism based on optimized perturbation noise needs to select different parameters for different data sets to balance the model performance and privacy protection, and better defense effect will correspondingly cause the model to perform significantly worse on the training task. The scheme based on autoencoder is mainly used to resist deep learning-based multi-classification label inference attacks, but the mapping space of its encoding is one-to-one mapping, and if the attacker knows a small amount of sample labels, the label privacy of the global data will also be leaked. Today, the dimension conversion mechanism is considered to be a convenient and effective scheme for label inference attack defense, but if the host party can perform clustering attacks (CA), such as using DBSCAN or K-Means clustering algorithm or other clustering algorithms to guess the elevated dimension parameters and using Model Completion Attack (MCA) to complete a top model on the bottom model of itself, and using a small number of labeled samples to train, after the parameter K is speculated, if the attacker also knows a part of the labels of the samples, then the dimension lifting / dimension reduction protection method can be cracked (here it is assumed that the attacker also has a part of the labels of the samples). Similarly, the label of the Guest party can be attacked to infer, and then the dimension lifting / dimension reduction protection scheme is cracked. In addition, if the labels of all tasks are converted in dimension, it may cause redundant calculation overhead.

[0106] In the present application, the scheme provided is a vertical federated deep learning scene of one Host party and one Guest party, wherein the label of the Guest party is M-classification (for example, M=2, [0, 1] represents a positive class, and [1, 0] represents a negative class) after one-hot encoding. A federated deep learning label protection scheme based on conditional dimension conversion and random noise is proposed.

[0107] As shown in the inference stage diagram Figure 4 In the federated model inference stage, the first participant inputs the first feature data X(A) into the first bottom model V(A), calculates the forward output of the first feature data through the first bottom model V(A), obtains the first forward output Z(A), and sends the first forward output Z(A) to the second participant. The second participant uses the locally trained second bottom model V(B) to calculate the forward output of the second feature data X(B), and obtains the second forward output Z(B). The second participant fuses the second forward output Z(B) and the first forward output Z(A) output by the first participant through a split layer to obtain a fusion result. The second participant processes the fusion result by using the trained top model T(B) to obtain a K-dimensional inference result If K is not equal to the preset dimension threshold parameter L, the K-dimensional inference result is taken as the final inference result.

[0108] If K is equal to the preset dimension threshold parameter L, which can be set according to actual conditions, for example, can be 3, 5, 8, etc. Then the decoded K-dimensional inference result is mapped to an M-dimensional inference result by using a dimension reduction module static mapping.

[0109] In the federated model inference stage, the Host uses the trained second bottom model to calculate the forward output (first forward output) of the local sample of the first feature data, and sends it to the Guest party. The Guest party also uses the locally trained second bottom model to calculate the forward output (second forward output) of the sample feature, and fuses the first forward output of the Host party to calculate The output of the dimension conversion is used as the static mapping The output of the dimension conversion is used as the static mapping The output of the dimension conversion is used as the static mapping The output of the dimension conversion is used as the static mapping The output of the dimension conversion is used as the static mapping

[0110] The output of the dimension conversion is used as the static mapping The output of the dimension conversion is used as the static mapping The output of the dimension conversion is used as the static mapping The output of the dimension conversion is used as the static mapping The output of the dimension conversion is used as the static mapping The output of the dimension conversion is used as the static mapping

[0111] The protection mechanism in the above embodiment is deployed on the Guest side. In a split learning scenario, if the Guest party does not have any feature data of the sample, only a top model needs to be trained. Therefore, the method has good scalability.

[0112] Optionally, after the first forward output and the second forward output are obtained, the method further includes: fusing the first forward output and the second forward output through a split layer in the second participant to obtain a fusion result; and processing the first forward output and the second forward output through a top model in the second participant to obtain a K-dimensional inference result, including: inputting the fusion result into the top model to obtain the K-dimensional inference result output by the top model.

[0113] As an optional implementation, the fusion manner of the split layer to the first forward output and the second forward output includes but is not limited to direct splicing, and can also be calculating a mean value, to obtain a fusion result, and inputting the fusion result into the top model T(B) to obtain the K-dimensional inference result output by the top model T(B). .

[0114] Optionally, the fusing the first forward output and the second forward output through the split layer in the second participant to obtain a fusion result includes: splicing the first forward output and the second forward output through the split layer in the second participant to obtain the fusion result; or performing mean value processing on the first forward output and the second forward output through the split layer in the second participant to obtain the fusion result.

[0115] As an optional implementation, the splicing can be directly splicing the first forward output and the second forward output, and the fusion result R after splicing is [Z(A), Z(B)]. Alternatively, the splicing can be mean processing of the first forward output and the second forward output, and the fusion result R after mean processing is (Z(A)+Z(B)) / 2.

[0116] Optionally, before the first forward output and the second forward output are obtained, the method further comprises: obtaining a first sample feature data set, a second sample feature data set, and an M-dimensional known sample label set having a corresponding relationship; performing sample label processing on each M-dimensional known sample label in the M-dimensional known sample label set to obtain a processed known sample label set, wherein the sample label processing comprises: adding random noise; and using the first sample feature data set, the second sample feature data set, and the processed known sample label set to jointly train a first bottom model to be trained in the first participant, a second bottom model to be trained in the second participant, and a top model to be trained until a first end condition is met, ending the training, and obtaining the first bottom model, the second bottom model, and the top model.

[0117] As an optional implementation, each M-dimensional known sample label in the M-dimensional known sample label set is used to represent an actual label, and the first sample feature data in the first sample feature data set and the second sample feature data in the second sample feature data set have the corresponding relationship.

[0118] Taking M=2 as an example above, in a vertical federated deep learning scenario of the first participant and the second participant, the M-dimensional known sample label provided by the second participant is a two-classification after one-hot encoding (for example, [0, 1] or [1, 0]), and the M-dimensional known sample label can be used to represent an actual sample operation result, which is used to represent whether a sample object feature represented by a sample object performs a target operation on a sample media resource feature represented by a sample media resource.

[0119] The first sample feature data in the first sample feature data set is a feature of the sample object. For example, the sample object is a sample user, and the first sample feature data is the gender, age, education, etc. of the sample user. The second sample feature data in the second sample feature data set can be a feature of the sample media resource, for example, a video category, a commodity type, etc. The first sample feature data, the second sample feature data and the M-dimensional known sample label in the first sample feature data set, the second sample feature data set and the M-dimensional known sample label set are corresponding. For example, the first sample feature data of the sample object is female, 15 years old, and a middle school student. The second sample feature data of the sample media resource is that the commodity type is Barbie doll, and the corresponding M-dimensional known sample label is [0, 1] (indicating that the actual sample operation result is purchase). For another example, the first sample feature data of the sample object is male, 55 years old, and the second sample feature data of the sample media resource is that the video content is a beauty blogger video. The corresponding M-dimensional known sample label is [1, 0], and the corresponding M-dimensional known sample label is [1, 0] (indicating that the actual sample operation result is not clicking to play).

[0120] In the above embodiment, there are two participants in the vertical federated deep learning scenario, wherein the first participant provides the first sample feature data, and the second participant provides the second sample feature data and the M-dimensional known sample label (the second participant can also only provide the M-dimensional known sample label). And the first participant and the second participant do not perform any sample feature and label data transmission. This scenario ensures that the sample data and the label data are only kept locally and will not leave the local, thereby providing privacy protection at the level of sample feature data and label data.

[0121] Optionally, the sample label processing on each M-dimensional known sample label in the M-dimensional known sample label set to obtain a processed known sample label set comprises: in the case that the M is less than the L, performing dimension lifting on the M-dimensional known sample label set by a dimension lifting module to obtain a K-dimensional known sample label set; adding random noise to each K-dimensional known sample label in the K-dimensional known sample label set to obtain a K-dimensional known noise sample label set, wherein the processed known sample label set comprises the K-dimensional known noise sample label set; or in the case that the M is greater than or equal to the L, adding random noise to each M-dimensional known sample label in the M-dimensional known sample label set to obtain an M-dimensional known noise sample label set, wherein the processed known sample label set comprises the M-dimensional known noise sample label set.

[0122] As an optional implementation, assuming that the Guest party is a M-classification task, the Guest party adds random noise to the one-hot encoded label vector to obtain a noise-perturbed classification label for federated learning model training, thereby avoiding attackers from inferring the parameter M.

[0123] Further, if the value of M is small, the Guest party can raise the dimension of the M-dimensional known sample label through a dimension raising module. Specifically, the Guest party self-determines a label dimension threshold parameter after dimension raising (which can be set according to actual conditions), if M < L, record K = L, and define a static mapping between M-dimensional and dimensional, respectively denoted as (from M-dimensional to dimensional) and (from dimensional to M-dimensional). The Guest party calculates the corresponding classification label of each sample for federated model training. If M >= L, there is no need to raise the label dimension. For ease of description, record K = M.

[0124] For example, when M = 2, L = 6, because M <= L, there can be:

[0125]

[0126]

[0127]

[0128] For the mapping from dimensional to M-dimensional, because the data in dimensional is not necessarily one-hot encoded, when M = 2, , based on the given , assuming that the inference result of the top model output is , a two-dimensional vector can be obtained through the following mapping:

[0129]

[0130] Where dot is the inner product operation of the vector. (Here, the conventional processing method can be to set the dimension with the largest value in to 1 and the rest to 0, and then map it back to M-dimensional. However, this method only considers one extreme dimension and ignores the values of other dimensions) If M >= L, there is no need to perform dimension conversion, record K = M. Then, the Guest party, based on the dimensional label set , for each label​ Randomly added noise.

[0131] In the federated training phase, the Host party uses the first base model to compute the forward output of the first sample feature data and sends it to the Guest party. The Guest party uses the second base model to compute the forward output of the local sample feature, and then uses the top model to compute the output after fusing the forward output of the Host party and the local forward output. Finally, the corresponding noise-perturbed classification label is used to compute the loss function and the gradient of backpropagation, updating the parameters of the top model and the base model.

[0132] For a given set of N samples , where is the sample feature set owned by the Host party (the first sample feature data set), is the sample feature set owned by the Guest party (the second sample feature data set), is the sample label set owned by the Guest party (the M-dimensional known sample label set).

[0133] Simply, we can use and to denote the features of a certain sample for the Host party and the Guest party, respectively, and use to denote the M classification label of the sample. For example, is used to represent the sample object feature, represents the media resource feature corresponding to the sample media resource, represents the actual sample operation result of the corresponding sample object pair the actual sample operation result of the corresponding sample media resource, such as purchasing goods, clicking to play a video, etc.

[0134] After the Guest party completes the conditional dimension conversion and adds noise to the M-dimensional known sample label, the Host party initializes a base model , the Guest party initializes a base model and a top model . At this time, for a given dimension , the sample set can be represented as . Simply, we can use and to denote the features of a certain sample for the Host party and the Guest party, respectively, and use to denote the noise-perturbed classification label of the sample.

[0135] The first sample feature data Input a first bottom model to be trained of a first participant, the first bottom model to be trained calculates a forward output first sample forward output , and the second sample feature data Input a second bottom model to be trained of a second participant, the second bottom model to be trained calculates a forward output second sample forward output , and the second participant's split layer and are fused (for example, can be directly spliced, or can be calculated as an average, etc.), to obtain a fusion result. The fusion result is input into a top model to be trained of the second participant to obtain a K-dimensional sample inference result output by the top model to be trained .

[0136] Then, a loss function L is calculated , wherein may be a cross-entropy loss function, is a K-dimensional known noise sample label, is a K-dimensional sample inference result output by the top model to be trained. If , the output value of the cross-entropy loss function between them is less than or equal to a preset threshold (the preset threshold can be determined according to actual conditions, for example, 0.01, 0.02, etc.), then the loss function converges.

[0137] The above first end condition includes but is not limited to: the loss function converges, the model parameters converge, the preset number of iterations is reached, the preset training duration is reached, etc. After the first end condition is met, the training is stopped, to obtain the first bottom model, the second bottom model, and the top model. If the above first end condition is not met, then the back propagation is performed, to modify the model parameters of the first bottom model to be trained, the second bottom model to be trained, and the top model to be trained.

[0138] Optionally, the sample label processing on each M-dimensional known sample label in the set of M-dimensional known sample labels includes: in the case where the M is less than the L, dimension increasing is performed on the set of M-dimensional known sample labels by a dimension increasing module, to obtain a set of K-dimensional known sample labels; random noise is added to each K-dimensional known sample label in the set of K-dimensional known sample labels, to obtain a set of K-dimensional known noise sample labels, wherein the set of processed known sample labels includes the set of K-dimensional known noise sample labels; or in the case where the M is greater than or equal to the L, random noise is added to each M-dimensional known sample label in the set of M-dimensional known sample labels, to obtain a set of M-dimensional known noise sample labels, wherein the set of processed known sample labels includes the set of M-dimensional known noise sample labels.

[0139] As an optional implementation, as shown in the flowchart, the method comprises the following steps: Figure 5

[0140] Step S501, input M-dimensional known sample labels;

[0141] Step S502, judge whether M is less than L, if M is less than L, execute step S503, if M is greater than or equal to L, execute step S504;

[0142] Step S503, dimension conversion, upgrade the M-dimensional known sample labels to K-dimensional, and obtain K-dimensional known sample labels;

[0143] Step S504, in the case that M is greater than or equal to L, add random noise to the M-dimensional known sample labels to obtain M-dimensional noise known sample labels; in the case that M is less than L, add random noise to the K-dimensional known sample labels to obtain K-dimensional known noise sample labels;

[0144] Step S505, train the bottom model and the top model of the Host party and the Gest party based on the noise-added sample labels (M-dimensional noise known sample labels or K-dimensional known noise sample labels);

[0145] Step S506, use the trained model to perform model inference;

[0146] Step S507, output the inference result.

[0147] The technical solution provided in the application is suitable for vertical federated deep learning. In an M-classification task, the Guest party of the vertical federated learning performs conditional dimension conversion and random noise addition locally. In the training phase, the Guest party performs federated deep learning model training based on the noise-added labels. In the inference phase, the Guest party also performs conditional dimension conversion to obtain an M-classification inference result.

[0148] ​Optionally, adding random noise to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain a K-dimensional known noise sample label set includes: adding randomly distributed noise to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set; or, adding noise of randomly opposite labels to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set; or, adding random opposite labels and randomly distributed noise to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set. The set; adding random noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain an M-dimensional known noise sample label set, including: adding randomly distributed noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain the M-dimensional known noise sample label set; or, adding noise of randomly opposite labels to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain the M-dimensional known noise sample label set; or, adding random opposite labels and randomly distributed noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain the M-dimensional known noise sample label set.

[0149] As an optional implementation, noise is randomly added to either the K-dimensional known sample labels or the M-dimensional noisy known sample labels. If the M-dimensional noisy known sample labels have been upsized (upgraded to K-dimensional known sample labels), then noise is randomly added to the resulting K-dimensional known sample labels. If the M-dimensional noisy known sample labels have not been upsized, then noise is randomly added to them. Methods for adding random noise include, but are not limited to: 1. noise based on random distribution, 2. noise based on random paired labels, and 3. noise based on both random paired labels and random distribution.

[0150] Optionally, adding randomly distributed noise to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set includes: performing the following steps on each of the K-dimensional known sample labels in the K-dimensional known sample label set, wherein each of the K-dimensional known sample labels during the execution of the following steps is the current K-dimensional known sample label: randomly generated through a specified random distribution. Dimensional random noise, and applied to the softmax function. The random noise of dimension is processed to obtain 3D random noise; The sum of the random noise and the current K-dimensional known sample label is determined as the K-dimensional known noise sample label of the current K-dimensional known sample label.

[0151] As an optional implementation, the Guest party adds noise to the K-dimensional known sample label set based on the dimension label set For each label Take random noise as an example:

[0152] For noise based on a random distribution, the Guest party adds noise to the K-dimensional known sample label set based on the dimension label set after dimension lifting For each K-dimensional known sample label Randomly generate a dimension random noise:

[0153] ,

[0154] Where the noise of each dimension is generated according to a certain specified random distribution, which can be a normal distribution, a Laplace distribution, etc.

[0155] After that, the size of the dimension random noise is limited using the softmax function, that is,

[0156] softmax( ) is used.

[0157] Subsequently, a corresponding random noise is added to each K-dimensional known sample label to obtain a K-dimensional known noise sample label, and the operation is as follows:

[0158]

[0159] Optionally, the K-dimensional known sample label set is added to each K-dimensional known sample label. The K-dimensional known noise sample label set is obtained by adding random noise to the opposite label, including: performing the following steps on each K-dimensional known sample label in the K-dimensional known sample label set, and each K-dimensional known sample label in the execution of the following steps is the current K-dimensional known sample label: obtaining the opposite label of the current K-dimensional known sample label, obtaining a K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; randomly select a K-dimensional opposite sample label from the K-dimensional opposite sample label set, and determine it as the dimension random noise of the current K-dimensional known sample label; and the sum of the dimension random noise and the current K-dimensional known sample label is determined as the K-dimensional known noise sample label of the current K-dimensional known sample label.

[0160] As an optional implementation, the Guest party is based on the upgraded dimension. Dimensional known sample label set For each known sample label Randomly generate one from its opposing sample labels. Dimensional random noise.

[0161] For example, suppose M = 2, If the label belongs to a negative sample, then randomly select one from its opposite sample label (positive sample label) as noise, as follows:

[0162] ;

[0163] Or assume If the label belongs to a positive sample, then randomly select one from its opposite sample label (negative sample label) as noise, as follows:

[0164] .

[0165] In short, the noise for each label is randomly drawn from the up-dimensional labels of its opposing sample labels.

[0166] Subsequently, labels were assigned to each K-dimensional known sample. Adding corresponding random noise to it yields a noisy result. K-dimensional labels (K-dimensional known noise sample labels), the operation is as follows:

[0167]

[0168] Optionally, adding random opposing labels and randomly distributed noise to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set includes: performing the following steps on each of the K-dimensional known sample labels in the K-dimensional known sample label set, where each of the K-dimensional known sample labels during the following steps is the current K-dimensional known sample label: randomly generating a random perturbation of the current K-dimensional known sample label through a specified random distribution; obtaining the opposing label of the current K-dimensional known sample label to obtain a K-dimensional opposing sample label set corresponding to the current K-dimensional known sample label; and determining the product of the randomly selected K-dimensional opposing sample label from the K-dimensional opposing sample label set and the random perturbation of the current K-dimensional known sample label as the current K-dimensional known sample label. 3D random noise; The sum of the random noise and the current K-dimensional known sample label is determined as the K-dimensional known noise sample label of the current K-dimensional known sample label.

[0169] As an optional implementation, for each noise , in the above embodiments, a random disturbance , is further introduced from a specified random distribution (normal distribution, Laplace distribution, etc.). The original is disturbed by , and the following is obtained:

[0170] .

[0171] Subsequently, a corresponding random noise is added to each K-dimensional known sample label , to obtain a noisy dimensional label, as follows:

[0172]

[0173] In the above embodiments, “ ” is vector addition, is an adjustable noise size coefficient. For example, when , , , then .

[0174] Optionally, the Guest party can perform softmax processing on , so that the sum of the elements of is 1.

[0175] Optionally, the joint training of the first bottom model to be trained in the first participant, the second bottom model to be trained in the second participant, and the top model to be trained, using the first sample feature data set, the second sample feature data set, and the processed known sample label set comprises: in the case that the processed known sample label set is the K-dimensional known noise sample label set, performing jthround of joint training on the first bottom model to be trained, the second bottom model to be trained, and the top model to be trained, where j is a positive integer greater than or equal to 1, and the first bottom model, the second bottom model, and the top model obtained through the 0thround of training are the first bottom model to be trained, the second bottom model to be trained, and the top model to be trained without training, and the jthround of joint training comprises: inputting S first sample feature data used in the jthround of training in the first sample feature data set into the first bottom model obtained through the (j-1) thround of training to obtain S first sample forward outputs outputted through the jthround of training, where S is a positive integer greater than or equal to 1; inputting S second sample feature data used in the jthround of training in the second sample feature data set into the second bottom model obtained through the (j-1) thround of training to obtain S second sample forward outputs outputted through the jthround of training; inputting the S first sample forward outputs outputted through the jthround of training and the S second sample forward outputs outputted through the jthround of training into the top model obtained through the (j-1) thround of training to obtain S K-dimensional sample inference results outputted through the jthround of training; and in the case that the S K-dimensional sample inference results outputted through the jthround of training and the S K-dimensional known noise sample labels used in the jthround of training satisfy the first end condition, ending the training to obtain the first bottom model, the second bottom model, and the top model, and the K-dimensional known noise sample labels used in the jthround of training have the corresponding relationship with the first sample feature data used in the jthround of training and the second sample feature data used in the jthround of training.

[0176] Optionally, in the case that the S K-dimensional sample inference results outputted through the jthround of training and the S K-dimensional known noise sample labels used in the jthround of training do not satisfy the first end condition, adjusting model parameters of the first bottom model obtained through the (j-1) thround of training, the second bottom model, and the top model obtained through the (j-1) thround of training through back propagation.

[0177] Optionally, the adjusting the model parameters of the top model obtained in the j-1th round of training through back propagation comprises: the second participant obtaining S first gradients through S loss functions between the S K-dimensional sample inference results and the S K-dimensional known noise sample labels and the S K-dimensional sample inference results, and adjusting the model parameters of the top model obtained in the j-1th round of training through the S first gradients; the adjusting the model parameters of the second bottom model obtained in the j-1th round of training through back propagation comprises: the second participant obtaining S second gradients through S loss functions between the S K-dimensional sample inference results and the S K-dimensional known noise sample labels and the model parameters of the second bottom model obtained in the j-1th round of training, and adjusting the model parameters of the second bottom model obtained in the j-1th round of training through the S second gradients; the adjusting the first bottom model obtained in the j-1th round of training through back propagation comprises: the second participant obtaining S third gradients through S loss functions between the S K-dimensional sample inference results and the S K-dimensional known noise sample labels and the S first sample forward outputs; the second participant sends the S third gradients to the first participant, and the first participant adjusts the first bottom model obtained in the j-1th round of training through the S third gradients.

[0178] As an optional implementation, as shown in FIG. 1, the model training process comprises the following steps. Figure 6 As shown in FIG. 2, which is a structural block diagram of model training, after the Guest completes the conditional dimension conversion and adds noise, the Host initializes a bottom model , the Guest initializes a bottom model and a top model . At this time, for a given dimension , the sample set can be represented as . Simply, the Host and the Guest can use and to respectively refer to the features of a sample, and use to represent the noise classification label of the sample. The Host calculates the forward output and sends it to the Guest. The Guest calculates the forward output and inputs the top model after fusing the received (the fusion method can be direct splicing or calculating the mean, etc.) to obtain the model output , and then calculates the loss function L= (taking the cross-entropy loss function as an example for illustration).

[0179] During the back propagation, the Guest party first calculates the gradient:

[0180] (second gradient)

[0181] and uses the Chain Rule to update the model , then the gradient calculation:

[0182] (first gradient)

[0183] for updating the model , here denotes the parameters of the model . Subsequently, the Guest party calculates the gradient:

[0184] (third gradient)

[0185] Subsequently, the Guest party sends to the Host party for parameter update of the bottom model . The above steps are performed once for each sample in each Epoch (iteration round), a total of Epochs are repeated until the model converges on the training sample set.

[0186] Optionally, the M-dimensional known sample label set is dimensionally upgraded by a dimension upgrading module, including: obtaining K K-dimensional one-hot encoding vectors, randomly dividing the K K-dimensional one-hot encoding vectors into M groups, or dividing the K K-dimensional one-hot encoding vectors into M groups according to a preset rule, to obtain M groups of K-dimensional one-hot encoding vectors; mapping the M-dimensional known sample labels to the M groups of K-dimensional one-hot encoding vectors respectively to obtain K-dimensional known sample labels corresponding to the M-dimensional known sample labels; and dimensionally reducing the K-dimensional inference result by a dimension reducing module in the second participant, including: performing inner product operation on the K-dimensional inference result and M K-dimensional decoding vectors to obtain the M-dimensional inference result, wherein the M K-dimensional decoding vectors are decoding vectors obtained according to the M groups of K-dimensional one-hot encoding.

[0187] As an optional implementation, taking M=2 above as an example, one-hot encoding can correspond to M categories, or can represent M+1 categories, including a full 0 vector (which can be used or not used), which can be used to represent whether an object performs a target operation on a media resource, such as a video click play operation, a product purchase operation, etc. It can be an integer multiple of M, or it can not be an integer multiple of M.

[0188] In the dimension upgrading phase, it is assumed that M=2, the dimension upgrading module is a static mapping from 2 dimensions to K dimensions may be:

[0189]

[0190] Firstly, K K-dimensional one-hot encoding vectors are generated, and in this embodiment, 4 4-dimensional one-hot encoding vectors are generated . The 4 4-dimensional one-hot encoding vectors are divided into M=2 groups randomly or according to a preset rule, where the random division can be that any one of the 4 4-dimensional one-hot encoding vectors is selected as a group, and the remaining three vectors are a group. The preset rule can be average allocation, that is, two of the 4 4-dimensional one-hot encoding vectors are a group, and the remaining two vectors are a group.

[0191] For example, the 4 4-dimensional one-hot encoding vectors are divided into and two groups. That is, the last two bits are added. Thus, the K-dimensional known sample label is obtained.

[0192] In the dimension reduction stage, based on the given , assuming that the decoded inference result is , a two-dimensional vector can be obtained through the following mapping:

[0193]

[0194] Where, dot is the inner product operation of the vector. The conventional processing method here is to set the dimension with the maximum value in to 1 and the remaining to 0, and then map it back to 2 dimensions. The above and are the corresponding decoding vectors.

[0195] When , M=2, the dimension increasing module is a static mapping from 2 dimensions to K dimensions may be:

[0196]

[0197] Firstly, K K-dimensional one-hot encoding vectors are generated, and in this embodiment, 6 6-dimensional one-hot encoding vectors are generated:

[0198]

[0199] The 6 six-dimensional one-hot encoding vectors are randomly or according to a preset rule divided into M=2 groups, wherein the random can be that any one or two vectors in the 6 six-dimensional one-hot encoding vectors are randomly selected as a group, and the remaining vectors are a group. The preset rule can be average allocation, that is, three vectors in the 6 six-dimensional one-hot encoding vectors are a group, and the remaining three vectors are a group.

[0200] Taking the case of dividing the 6 six-dimensional one-hot encoding vectors into two groups by average allocation as an example.

[0201]

[0202]

[0203] In , the sample object performs the target operation on the sample media resource, , the sample object performs the target operation on the sample media resource. Similarly, in , the sample object does not perform the target operation on the sample media resource, , the sample object does not perform the target operation on the sample media resource.

[0204] Regarding , the mapping from 6 dimensions to 2 dimensions, because , the data in the 6 dimensions is not necessarily one-hot encoding, when , based on the given , assuming that the decoded inference result is , a two-dimensional vector can be obtained through the following mapping

[0205]

[0206] wherein dot is the inner product operation of the vector, and the conventional processing method is to set the dimension with the largest value in to 1 and the remaining to 0, and then map back to 2 dimensions.

[0207] Suppose ,

[0208]

[0209] In the above embodiment, the sample feature data of the first participant (Host party) and the second participant (Guest party) are all in the local, and the second participant realizes the label confusion and the protection of the dimension information of the second participant through the combination of the label conditional dimension conversion and the random noise. It should be noted that if in the traditional vertical federated learning scenario, that is, the second participant has a certain number of features, the above embodiment can also be directly executed, because the label confusion module does not affect the performance of the training task.

[0210] As an optional implementation, a federal deep model training process is referenced Figure 7 , and a federal deep model usage process is referenced Figure 8 .

[0211] In Figure 7 the federal deep model training process includes the following steps:

[0212] The following steps are performed for the first participant:

[0213] Step S701, obtaining first sample feature data in the first sample feature data set, the first sample feature data can be the feature of the target object;

[0214] Step S702, inputting the first sample feature data into the first bottom model to be trained of the first participant to obtain first sample forward output;

[0215] The following steps are performed for the second participant:

[0216] Step S703, obtaining second sample feature data and M-dimensional known sample labels in the second sample feature data set, the second sample feature data can be the feature of the target media resource;

[0217] Step S704, inputting the second sample feature data into the second bottom model to be trained of the second participant to obtain second sample forward output;

[0218] Step S705, determining whether M is less than L, if M is less than L, performing step S706, otherwise performing step S707;

[0219] Step S706, inputting the M-dimensional known sample label into the dimension increasing module of the second participant to obtain the K-dimensional known sample label;

[0220] Step S707, adding noise to the K-dimensional known sample label or the M-dimensional known sample label to obtain the K-dimensional known noise sample label or the M-dimensional known noise sample label;

[0221] Step S708, fusing the first sample forward output and the second sample forward output through the splitting layer of the second participant to obtain a fusion result;

[0222] Step S709, processing the fusion result through the top model of the second participant to obtain a K-dimensional sample inference result;

[0223] Step S710, determining whether the K-dimensional sample inference result and the K-dimensional known noise sample label (or the M-dimensional known noise sample label) satisfy a first end condition, and ending the training to obtain the first bottom model, the second bottom model and the top model if the first end condition is satisfied, and adjusting the model parameters and continuing the training if the first end condition is not satisfied.

[0224] In Figure 8 The federated deep model uses the following steps:

[0225] The following steps are performed for the first participant:

[0226] Step S801, obtaining first feature data, such as gender, age, etc. of a target object;

[0227] Step S802, inputting the first feature data into the first bottom model of the first participant to obtain a first forward output;

[0228] The following steps are performed for the second participant:

[0229] Step S803, obtaining second feature data, such as video content, product type, etc.;

[0230] Step S804, inputting the second feature data into the second bottom model of the second participant to obtain a second forward output;

[0231] Step S805, fusing the first forward output and the second forward output through the split layer of the second participant to obtain a fusion result;

[0232] Step S806, processing the fusion result through the top model of the second participant to obtain a K-dimensional inference result;

[0233] Step S807, if K is equal to L, dimensionally reducing the K-dimensional inference result through the dimension reduction module to obtain an M-dimensional inference result. If K is not equal to L, taking the K-dimensional inference result as the final inference result.

[0234] According to another aspect of the embodiments of the present application, a data processing method is also provided, which is a training process of the first bottom model in the first participant, the second bottom model in the second participant and the top model, and the training manner of the model is the same as that of the first bottom model, the second bottom model and the top model in the above embodiments. The data processing method comprises:

[0235] Step S902, obtaining a first sample feature data set, a second sample feature data set and an M-dimensional known sample label set having a corresponding relationship, wherein each M-dimensional known sample label in the M-dimensional known sample label set is used to represent an actual sample label;

[0236] In step S904, sample label processing is performed on each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain a set of processed known sample labels, wherein the sample label processing comprises adding random noise.

[0237] In step S906, the set of first sample feature data, the set of second sample feature data, and the set of processed known sample labels are used to jointly train the first bottom model in the first participant, the second bottom model in the second participant, and the top model until a first end condition is met, and the training is ended to obtain the first bottom model, the second bottom model, and the top model.

[0238] Optionally, the sample label processing performed on each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain a set of processed known sample labels comprises: in a case where the M is less than the L, dimension increasing is performed on the set of M-dimensional known sample labels by a dimension increasing module to obtain a set of K-dimensional known sample labels; random noise is added to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain a set of K-dimensional known noise sample labels, wherein the set of processed known sample labels comprises the set of K-dimensional known noise sample labels; or in a case where the M is greater than or equal to the L, random noise is added to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain a set of M-dimensional known noise sample labels, wherein the set of processed known sample labels comprises the set of M-dimensional known noise sample labels.

[0239] Optionally, adding random noise to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain a K-dimensional known noise sample label set includes: adding randomly distributed noise to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set; or, adding noise of randomly opposite labels to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set; or, adding random opposite labels and randomly distributed noise to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set. The set; adding random noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain an M-dimensional known noise sample label set, including: adding randomly distributed noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain the M-dimensional known noise sample label set; or, adding noise of randomly opposite labels to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain the M-dimensional known noise sample label set; or, adding random opposite labels and randomly distributed noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain the M-dimensional known noise sample label set.

[0240] Optionally, adding randomly distributed noise to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set includes: performing the following steps on each of the K-dimensional known sample labels in the K-dimensional known sample label set, wherein each of the K-dimensional known sample labels during the execution of the following steps is the current K-dimensional known sample label: randomly generated through a specified random distribution. Dimensional random noise, and applied to the softmax function. The random noise of dimension is processed to obtain 3D random noise; The sum of the random noise and the current K-dimensional known sample label is determined as the K-dimensional known noise sample label of the current K-dimensional known sample label.

[0241] Optionally, the K-dimensional known sample label set is added with random opposite label noise to obtain the K-dimensional known noise sample label set, including: performing the following steps on each K-dimensional known sample label in the K-dimensional known sample label set, and each K-dimensional known sample label in the performing of the following steps is a current K-dimensional known sample label: obtaining the opposite label of the current K-dimensional known sample label to obtain a K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; determining the K-dimensional opposite sample label randomly selected from the K-dimensional opposite sample label set as the K-dimensional known noise sample label of the current K-dimensional known sample label. Optionally, the K-dimensional known sample label set is added with random opposite label noise to obtain the K-dimensional known noise sample label set, including: performing the following steps on each K-dimensional known sample label in the K-dimensional known sample label set, and each K-dimensional known sample label in the performing of the following steps is a current K-dimensional known sample label: obtaining the opposite label of the current K-dimensional known sample label to obtain a K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; determining the K-dimensional opposite sample label randomly selected from the K-dimensional opposite sample label set as the K-dimensional known noise sample label of the current K-dimensional known sample label. Optionally, the K-dimensional known sample label set is added with random opposite label noise to obtain the K-dimensional known noise sample label set, including: performing the following steps on each K-dimensional known sample label in the K-dimensional known sample label set, and each K-dimensional known sample label in the performing of the following steps is a current K-dimensional known sample label: obtaining the opposite label of the current K-dimensional known sample label to obtain a K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; determining the K-dimensional opposite sample label randomly selected from the K-dimensional opposite sample label set as the K-dimensional known noise sample label of the current K-dimensional known sample label.

[0242] Optionally, the K-dimensional known sample label set is added with random opposite label noise to obtain the K-dimensional known noise sample label set, including: performing the following steps on each K-dimensional known sample label in the K-dimensional known sample label set, and each K-dimensional known sample label in the performing of the following steps is a current K-dimensional known sample label: obtaining the opposite label of the current K-dimensional known sample label to obtain a K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; determining the K-dimensional opposite sample label randomly selected from the K-dimensional opposite sample label set as the K-dimensional known noise sample label of the current K-dimensional known sample label. Optionally, the K-dimensional known sample label set is added with random opposite label noise to obtain the K-dimensional known noise sample label set, including: performing the following steps on each K-dimensional known sample label in the K-dimensional known sample label set, and each K-dimensional known sample label in the performing of the following steps is a current K-dimensional known sample label: obtaining the opposite label of the current K-dimensional known sample label to obtain a K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; determining the K-dimensional opposite sample label randomly selected from the K-dimensional opposite sample label set as the K-dimensional known noise sample label of the current K-dimensional known sample label. Optionally, the K-dimensional known sample label set is added with random opposite label noise to obtain the K-dimensional known noise sample label set, including: performing the following steps on each K-dimensional known sample label in the K-dimensional known sample label set, and each K-dimensional known sample label in the performing of the following steps is a current K-dimensional known sample label: obtaining the opposite label of the current K-dimensional known sample label to obtain a K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; determining the K-dimensional opposite sample label randomly selected from the K-dimensional opposite sample label set as the K-dimensional known noise sample label of the current K-dimensional known sample label.

[0243] Optionally, the K-dimensional known sample label set is added with random opposite label noise to obtain the K-dimensional known noise sample label set, including: performing the following steps on each K-dimensional known sample label in the K-dimensional known sample label set, and each K-dimensional known sample label in the performing of the following steps is a current K-dimensional known sample label: obtaining the opposite label of the current K-dimensional known sample label to obtain a K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; determining the K-dimensional opposite sample label randomly selected from the K-dimensional opposite sample label set as the K-dimensional known noise sample label of the current K-dimensional known sample label.

[0244] In a case that the processed known sample label set is the K-dimensional known noise sample label set, the first to-be-trained bottom model, the second to-be-trained bottom model and the to-be-trained top model are trained for the jth round, where j is a positive integer greater than or equal to 1, and the first bottom model, the second bottom model and the top model obtained by the 0th round of training are the to-be-trained first bottom model, the to-be-trained second bottom model and the to-be-trained top model, and the training comprises the following steps:

[0245] S first sample feature data used in the jth round of the first sample feature data set are input into the first bottom model trained in the (j-1)th round to obtain S first sample forward outputs output by the jth round of training, where S is a positive integer greater than or equal to 1; S second sample feature data used in the jth round of the second sample feature data set are input into the second bottom model trained in the (j-1)th round to obtain S second sample forward outputs output by the jth round of training; the S first sample forward outputs output by the jth round of training and the S second sample forward outputs output by the jth round of training are input into the top model trained in the (j-1)th round to obtain S K-dimensional sample inference results output by the jth round of training; in a case that the first end condition is met between the S K-dimensional sample inference results output by the jth round of training and the S K-dimensional known noise sample labels used in the jth round, the training is ended to obtain the first bottom model, the second bottom model and the top model, and the K-dimensional known noise sample labels used in the jth round have the corresponding relationship with the first sample feature data used in the jth round and the second sample feature data used in the jth round.

[0246] Optionally, the method further comprises: in a case that the first end condition is not met between the S K-dimensional sample inference results output by the jth round of training and the S K-dimensional known noise sample labels used in the jth round, adjusting model parameters of the first bottom model trained in the (j-1)th round, the second bottom model and the top model trained in the (j-1)th round by back propagation.

[0247] Optionally, the adjusting of the model parameters of the top model trained in the (j-1)th round by back propagation comprises: the second participant obtains S first gradients from S loss functions between the S K-dimensional sample inference results and the S K-dimensional known noise sample labels and the S K-dimensional sample inference results, and adjusts the model parameters of the top model trained in the (j-1)th round by the S first gradients.

[0248] The adjusting the model parameters of the second base model obtained by the j-1th round of training through back propagation comprises: the second participant obtaining S second gradients through S loss functions between the S K-dimensional sample inference results and the S K-dimensional known noise sample labels and the model parameters of the second base model obtained by the j-1th round of training, and adjusting the model parameters of the second base model obtained by the j-1th round of training through the S second gradients.

[0249] The adjusting the first base model obtained by the j-1th round of training through back propagation comprises: the second participant obtaining S third gradients through S loss functions between the S K-dimensional sample inference results and the S K-dimensional known noise sample labels and the S first sample forward outputs; and the second participant sending the S third gradients to the first participant, and the first participant adjusting the first base model obtained by the j-1th round of training through the S third gradients.

[0250] Optionally, the dimension increasing module is configured to increase dimensions of the M-dimensional known sample label set, comprising: obtaining K K-dimensional one-hot encoding vectors, randomly dividing the K K-dimensional one-hot encoding vectors into M groups, or dividing the K K-dimensional one-hot encoding vectors into M groups according to a preset rule, to obtain M groups of K-dimensional one-hot encoding vectors; and mapping the M-dimensional known sample labels to the M groups of K-dimensional one-hot encoding vectors respectively to obtain K-dimensional known sample labels corresponding to the M-dimensional known sample labels.

[0251] In the above embodiment, in the existing noise-based label protection scheme, a random or optimized disturbance is added to the gradient of the Guest direction back propagation to make the gradients corresponding to different categories of sample labels indistinguishable. The isotropic Gaussian noise-based scheme adds noise different from the original gradient distribution to the original back propagation gradient, which will cause the performance of the model on the training task to decline. The optimized disturbance noise-based scheme adds different noise to the back propagation gradient of different labels, but different hyperparameters need to be adjusted for different data sets, and better defense effect will correspondingly cause the performance of the model on the training task to decline significantly. In the existing self-encoder-based label protection scheme, the Guest trains an encoder to encode the sample labels in the same dimension, and uses a decoder to decode the inference result in the final inference stage. However, the mapping space of the encoding is one-to-one mapping, and if the attacker knows a small amount of sample labels, the global data label privacy will also be leaked. In the existing label protection scheme based on single dimension conversion mechanism, if the Host performs clustering attack, for example, uses K-Means clustering algorithm to speculate the increased dimension And using a model completion attack to complete a top model on the bottom model of itself, and using a small amount of labeled samples to train, can also perform a label inference attack on the Guest side.

[0252] The technical solution provided in the present application provides a label-protected longitudinal federated deep learning method based on conditional dimension conversion and random noise. In the technical solution provided in the present application, the Guest side uses conditional dimension conversion to reduce the dimension conversion overhead of the classification task, expand the label encoding space of the classification task, obtain K classification labels and add random noise, without adding random or optimized noise to the back propagation gradient. After adding noise, the Guest side can directly transmit the back propagation gradient obtained by training to the Host side, without leaking the dimension of the training label , so that clustering attacks and model completion attacks cannot be implemented, thereby avoiding the risk of leaking the real label of the sample. In the technical solution provided in the present application, the label protection method based on random noise perturbation can be applied to any classification data set, and can reduce the dimension conversion overhead and solve the problem that the traditional dimension conversion scheme is easy to be inferred by a clustering attack to increase the dimension .

[0253] The technical solution provided in the present application is applicable to a longitudinal federated deep learning scenario, and can cope with clustering attacks, model completion attacks and label inference attacks that the Host side may implement in the scenario. The present solution mainly protects the model training under a classification task. The functional performance of the longitudinal federated deep learning product side can include a model training module as a federated learning task, improve the usability of the federated learning system platform, and increase the functional module of the federated learning system, i.e. increase the label protection scheme based on label dimension conversion and random noise, and improve the privacy and security characteristics of the federated learning model. The main product form can be in the form of a federated learning platform to provide federated learning services to the public cloud or private cloud. There is no mature product on the current market, only related open source software, including FATE open sourced by Shenzhen Qianhai Micro-World Bank, PaddleFL open sourced by Baidu, Fedlearner open sourced by ByteDance, TensorFlowFederated (TFF) open sourced by Google, and PySyft open sourced by OpenMined. However, these open source software do not support the federated deep learning label protection scheme based on label dimension conversion and random noise.

[0254] In the technical solution provided in the present application, the Guest side uses conditional label dimension conversion and random noise to protect its sample label data and dimension information , Host party does not need to carry out any additional design and operation. The classification task can be carried out by the Host party, and the clustering attack, model completion attack and label stealing attack can be carried out. The multi-classification label information can be confused by random noise, and the clustering attack and model completion attack that the traditional dimension conversion scheme may suffer can be solved. The traditional longitudinal federated deep learning and split learning scene can be supported.

[0255] In the technical scheme provided in the present application, all protection mechanisms are carried out at the Guest end. In the split learning scene, the Guest party can not have any feature data of the sample, and only a top model needs to be trained. Therefore, the present scheme has good scalability and can also be applied to the split learning scene. Based on this, according to another aspect of the embodiment of the present application, a data processing method is also provided, which is different from the above-mentioned embodiment in that the Guest party (the second participant) does not provide the second feature data, and the second bottom model of the Guest party (the second participant) does not need to be trained, and the Guest party (the second participant) only needs to train the top model. The method comprises:

[0256] Step S1002, obtaining a first forward output, wherein the first forward output is an output obtained by processing first feature data by a first bottom model in a first participant;

[0257] Step S1004, processing the first forward output by a top model in the second participant to obtain a K-dimensional inference result, wherein K is a positive integer greater than 2;

[0258] Step S1006, in the case where K is equal to L, reducing the dimension of the K-dimensional inference result by a dimension reduction module in the second participant to obtain an M-dimensional inference result, wherein M is a positive integer less than K and greater than or equal to 2, and L is a preset dimension threshold parameter; wherein the first bottom model and the top model are obtained by jointly training a to-be-trained first bottom model in the first participant and a to-be-trained top model in the second participant using a first sample feature data set and a processed known sample label set, the processed known sample label set is obtained by processing each M-dimensional known sample label in an M-dimensional known sample label set, and the sample label processing includes adding random noise.

[0259] As an optional implementation, the embodiment is different from the above-mentioned embodiments in that the second participant (Guest) does not provide the second feature data in the embodiment. In the model inference stage, the input of the model is the first feature data provided by the first participant (Host party), which can be the feature data of a target object, which can be a target user, and the first feature data can be the portrait features of the target user (for example, gender, age, education, location, etc.), or can be the feature data of a target media resource, which includes but is not limited to video, audio, commodity, etc. media resources.

[0260] The first feature data is input into the trained first bottom model to obtain a first forward output output by the first bottom model, and the first forward output is processed by the trained top model in the second participant to obtain a K-dimensional inference result, which can be used to represent the classification result of the first feature data, such as the classification result of the above-mentioned target user, or the classification result of the above-mentioned video, audio, etc.

[0261] If K is not equal to the preset dimension threshold parameter L, the K-dimensional inference result is as the final inference result. If it is equal to the preset dimension threshold parameter L, it can be set according to actual conditions, for example, it can be 3, 5, 8, etc. Then use the dimension reduction module static mapping to map the decoded K-dimensional inference result to the M-dimensional inference result.

[0262] Optionally, before the first forward output is obtained, the method further comprises: obtaining a first sample feature data set with a corresponding relationship, and an M-dimensional known sample label set, wherein each M-dimensional known sample label in the M-dimensional known sample label set is used to represent an actual sample label; performing sample label processing on each M-dimensional known sample label in the M-dimensional known sample label set to obtain a processed known sample label set, wherein the sample label processing comprises: adding random noise; using the first sample feature data set and the processed known sample label set to jointly train the to-be-trained first bottom model in the first participant and the to-be-trained top model in the second participant until a first end condition is met, and ending the training to obtain the first bottom model and the top model.

[0263] As an optional implementation, in the model training stage, the embodiment is different from the above-mentioned embodiments in that the second bottom model in the second participant does not need to be trained in the embodiment, and the second participant does not need to provide the second sample feature data set.

[0264] The first sample feature data can be feature data of a sample user, such as age, gender, and the like of the sample user, or can be feature data of a video, audio, and the like. The M-dimensional known sample label is used to represent an actual label, and can be used to represent an actual classification result of the first sample feature data, for example, an actual classification result of a video, audio, and the like, which is known.

[0265] In this embodiment, the first bottom model in the first participant and the top model in the second participant are trained by the first sample feature data set provided by the first participant and the M-dimensional known sample label set provided by the second participant. For details, refer to the above embodiment.

[0266] Optionally, the sample label processing on each of the M-dimensional known sample labels in the M-dimensional known sample label set to obtain a processed known sample label set comprises: in the case that the M is less than the L, dimension increasing is performed on the M-dimensional known sample label set by a dimension increasing module to obtain a K-dimensional known sample label set; random noise is added to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain a K-dimensional known noise sample label set, wherein the processed known sample label set comprises the K-dimensional known noise sample label set; or in the case that the M is greater than or equal to the L, random noise is added to each of the M-dimensional known sample labels in the M-dimensional known sample label set to obtain an M-dimensional known noise sample label set, wherein the processed known sample label set comprises the M-dimensional known noise sample label set.

[0267] Optionally, adding random noise to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain a K-dimensional known noise sample label set includes: adding randomly distributed noise to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set; or, adding noise of randomly opposite labels to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set; or, adding random opposite labels and randomly distributed noise to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set. The set; adding random noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain an M-dimensional known noise sample label set, including: adding randomly distributed noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain the M-dimensional known noise sample label set; or, adding noise of randomly opposite labels to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain the M-dimensional known noise sample label set; or, adding random opposite labels and randomly distributed noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain the M-dimensional known noise sample label set.

[0268] Optionally, adding randomly distributed noise to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set includes: performing the following steps on each of the K-dimensional known sample labels in the K-dimensional known sample label set, wherein each of the K-dimensional known sample labels during the execution of the following steps is the current K-dimensional known sample label: randomly generated through a specified random distribution. Dimensional random noise, and applied to the softmax function. The random noise of dimension is processed to obtain 3D random noise; The sum of the random noise and the current K-dimensional known sample label is determined as the K-dimensional known noise sample label of the current K-dimensional known sample label.

[0269] Optionally, adding random opposite label noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels comprises: performing the following steps on each of the K-dimensional known sample labels in the set of K-dimensional known sample labels, each of the K-dimensional known sample labels when performing the following steps being a current K-dimensional known sample label: obtaining the opposite label of the current K-dimensional known sample label to obtain a set of K-dimensional opposite sample labels corresponding to the current K-dimensional known sample label; determining a K-dimensional opposite sample label randomly selected from the set of K-dimensional opposite sample labels as the K-dimensional known noise sample label of the current K-dimensional known sample label. Optionally, adding random opposite label noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels comprises: performing the following steps on each of the K-dimensional known sample labels in the set of K-dimensional known sample labels, each of the K-dimensional known sample labels when performing the following steps being a current K-dimensional known sample label: obtaining the opposite label of the current K-dimensional known sample label to obtain a set of K-dimensional opposite sample labels corresponding to the current K-dimensional known sample label; determining a K-dimensional opposite sample label randomly selected from the set of K-dimensional opposite sample labels as the K-dimensional known noise sample label of the current K-dimensional known sample label. Optionally, adding random opposite label noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels comprises: performing the following steps on each of the K-dimensional known sample labels in the set of K-dimensional known sample labels, each of the K-dimensional known sample labels when performing the following steps being a current K-dimensional known sample label: obtaining the opposite label of the current K-dimensional known sample label to obtain a set of K-dimensional opposite sample labels corresponding to the current K-dimensional known sample label; determining a K-dimensional opposite sample label randomly selected from the set of K-dimensional opposite sample labels as the K-dimensional known noise sample label of the current K-dimensional known sample label.

[0270] Optionally, adding random opposite label noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels comprises: performing the following steps on each of the K-dimensional known sample labels in the set of K-dimensional known sample labels, each of the K-dimensional known sample labels when performing the following steps being a current K-dimensional known sample label: obtaining the opposite label of the current K-dimensional known sample label to obtain a set of K-dimensional opposite sample labels corresponding to the current K-dimensional known sample label; determining a K-dimensional opposite sample label randomly selected from the set of K-dimensional opposite sample labels as the K-dimensional known noise sample label of the current K-dimensional known sample label. Optionally, adding random opposite label noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels comprises: performing the following steps on each of the K-dimensional known sample labels in the set of K-dimensional known sample labels, each of the K-dimensional known sample labels when performing the following steps being a current K-dimensional known sample label: obtaining the opposite label of the current K-dimensional known sample label to obtain a set of K-dimensional opposite sample labels corresponding to the current K-dimensional known sample label; determining a K-dimensional opposite sample label randomly selected from the set of K-dimensional opposite sample labels as the K-dimensional known noise sample label of the current K-dimensional known sample label. Optionally, adding random opposite label noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels comprises: performing the following steps on each of the K-dimensional known sample labels in the set of K-dimensional known sample labels, each of the K-dimensional known sample labels when performing the following steps being a current K-dimensional known sample label: obtaining the opposite label of the current K-dimensional known sample label to obtain a set of K-dimensional opposite sample labels corresponding to the current K-dimensional known sample label; determining a K-dimensional opposite sample label randomly selected from the set of K-dimensional opposite sample labels as the K-dimensional known noise sample label of the current K-dimensional known sample label.

[0271] Optionally, the joint training of the first bottom model to be trained in the first participant and the top model to be trained in the second participant using the first sample feature data set and the processed known sample label set comprises: in a case where the processed known sample label set is the K-dimensional known noise sample label set, performing j-th round of joint training on the first bottom model to be trained and the top model to be trained, where j is a positive integer greater than or equal to 1, and the first bottom model and the top model obtained through 0-th round of training are the first bottom model to be trained and the top model to be trained without training, and the j-th round of joint training comprises: inputting S first sample feature data used in the j-th round of training in the first sample feature data set into the first bottom model obtained through (j-1)-th round of training to obtain S first sample forward outputs outputted through the j-th round of training, where S is a positive integer greater than or equal to 1; inputting the S first sample forward outputs outputted through the j-th round of training into the top model obtained through (j-1)-th round of training to obtain S K-dimensional sample inference results outputted through the j-th round of training; and in a case where the S K-dimensional sample inference results outputted through the j-th round of training and S K-dimensional known noise sample labels used in the j-th round of training satisfy the first end condition, ending the training to obtain the first bottom model and the top model, and the K-dimensional known noise sample labels used in the j-th round of training have the corresponding relationship with the first sample feature data used in the j-th round of training.

[0272] Optionally, the method further comprises: in a case where the S K-dimensional sample inference results outputted through the j-th round of training and the S K-dimensional known noise sample labels used in the j-th round of training do not satisfy the first end condition, adjusting model parameters of the first bottom model obtained through the (j-1)-th round of training and the top model obtained through the (j-1)-th round of training through back propagation.

[0273] Optionally, the adjusting of the model parameters of the top model obtained through the (j-1)-th round of training through back propagation comprises: obtaining, by the second participant, S first gradients through S loss functions between the S K-dimensional sample inference results and the S K-dimensional known noise sample labels and the S K-dimensional sample inference results, and adjusting the model parameters of the top model obtained through the (j-1)-th round of training through the S first gradients.

[0274] The adjusting the first bottom model obtained through the j-1th round of training by back propagation comprises: the second participant obtaining S third gradients through S loss functions between the S K-dimensional inference results and the S K-dimensional known noise sample labels and the S first sample forward outputs; and the second participant sending the S third gradients to the first participant, and the first participant adjusting the first bottom model obtained through the j-1th round of training through the S third gradients.

[0275] Optionally, the M-dimensional known sample label set is dimensionally upgraded through a dimension upgrading module, comprising: obtaining K K-dimensional one-hot encoding vectors, randomly dividing the K K-dimensional one-hot encoding vectors into M groups, or dividing the K K-dimensional one-hot encoding vectors into M groups according to a preset rule, to obtain M groups of K-dimensional one-hot encoding vectors; and mapping the M-dimensional known sample labels to the M groups of K-dimensional one-hot encoding vectors respectively, to obtain K-dimensional known sample labels corresponding to the M-dimensional known sample labels.

[0276] The K-dimensional inference result is dimensionally reduced through a dimension reducing module in the second participant, comprising: performing inner product operation on the K-dimensional inference result and M K-dimensional decoding vectors to obtain the M-dimensional inference result, wherein the M K-dimensional decoding vectors are decoding vectors obtained according to the M groups of K-dimensional one-hot encoding.

[0277] According to another aspect of the embodiments of the present application, a data processing method is further provided, which is a training method for a first bottom model of a first participant and a top model of a second participant, and the method comprises:

[0278] In step S1102, a first sample feature data set and an M-dimensional known sample label set having a corresponding relationship are obtained, wherein each M-dimensional known sample label in the M-dimensional known sample label set is used to represent an actual sample label.

[0279] In step S1104, each M-dimensional known sample label in the M-dimensional known sample label set is subjected to sample label processing to obtain a processed known sample label set, wherein the sample label processing comprises adding random noise.

[0280] In step S1106, the first sample feature data set and the processed known sample label set are used to jointly train a to-be-trained first bottom model in the first participant and a to-be-trained top model in the second participant until a first end condition is met, and the training is ended to obtain the first bottom model and the top model.

[0281] Optionally, the sample label processing on each of the M-dimensional known sample labels in the set of M-dimensional known sample labels comprises: in the case that the M is less than the L, dimension increasing on the set of M-dimensional known sample labels by a dimension increasing module to obtain a set of K-dimensional known sample labels; adding random noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain a set of K-dimensional known noise sample labels, wherein the set of processed known sample labels comprises the set of K-dimensional known noise sample labels; or in the case that the M is greater than or equal to the L, adding random noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain a set of M-dimensional known noise sample labels, wherein the set of processed known sample labels comprises the set of M-dimensional known noise sample labels.

[0282] Optionally, the adding random noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain a set of K-dimensional known noise sample labels comprises: adding randomly distributed noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels; or adding randomly opposite label noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels; or adding randomly opposite label and randomly distributed noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels; the adding random noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain a set of M-dimensional known noise sample labels comprises: adding randomly distributed noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain the set of M-dimensional known noise sample labels; or adding randomly opposite label noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain the set of M-dimensional known noise sample labels; or adding randomly opposite label and randomly distributed noise to each of the M-dimensional known sample labels in the set of M-dimensional known sample labels to obtain the set of M-dimensional known noise sample labels.

[0283] Optionally, the adding randomly distributed noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels comprises: performing the following steps on each of the K-dimensional known sample labels in the set of K-dimensional known sample labels, wherein each of the K-dimensional known sample labels in the performing of the following steps is a current K-dimensional known sample label: generating a random noise of the K dimension by a specified random distribution, and processing the random noise of the K dimension by a softmax function to obtain a random noise of the L dimension. Optionally, the adding randomly distributed noise to each of the K-dimensional known sample labels in the set of K-dimensional known sample labels to obtain the set of K-dimensional known noise sample labels comprises: performing the following steps on each of the K-dimensional known sample labels in the set of K-dimensional known sample labels, wherein each of the K-dimensional known sample labels in the performing of the following steps is a current K-dimensional known sample label: generating a random noise of the K dimension by a specified random distribution, and processing the random noise of the K dimension by a softmax function to obtain a random noise of the L dimension. a K-dimensional random noise of the K-dimensional known sample label is added to the K-dimensional known sample label to obtain a K-dimensional known noise sample label of the K-dimensional known sample label. a K-dimensional random noise of the K-dimensional known sample label is added to the K-dimensional known sample label to obtain a K-dimensional known noise sample label of the K-dimensional known sample label. a K-dimensional random noise of the K-dimensional known sample label is added to the K-dimensional known sample label to obtain a K-dimensional known noise sample label of the K-dimensional known sample label.

[0284] Optionally, a random noise of an opposite label is added to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set, including: performing the following steps on each of the K-dimensional known sample labels in the K-dimensional known sample label set, and each of the K-dimensional known sample labels in the performing of the following steps is a current K-dimensional known sample label: obtaining an opposite label of the current K-dimensional known sample label to obtain a K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; and determining a K-dimensional opposite sample label randomly selected from the K-dimensional opposite sample label set as a K-dimensional random noise of the current K-dimensional known sample label. a K-dimensional random noise of the K-dimensional known sample label is added to the K-dimensional known sample label to obtain a K-dimensional known noise sample label of the K-dimensional known sample label. a K-dimensional random noise of the K-dimensional known sample label is added to the K-dimensional known sample label to obtain a K-dimensional known noise sample label of the K-dimensional known sample label.

[0285] Optionally, a random noise of an opposite label and a random distribution noise are added to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set, including: performing the following steps on each of the K-dimensional known sample labels in the K-dimensional known sample label set, and each of the K-dimensional known sample labels in the performing of the following steps is a current K-dimensional known sample label: randomly generating a random disturbance of the current K-dimensional known sample label through a specified random distribution; obtaining an opposite label of the current K-dimensional known sample label to obtain a K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; and determining a product of a K-dimensional opposite sample label randomly selected from the K-dimensional opposite sample label set and the random disturbance of the current K-dimensional known sample label as a K-dimensional random noise of the current K-dimensional known sample label. a K-dimensional random noise of the K-dimensional known sample label is added to the K-dimensional known sample label to obtain a K-dimensional known noise sample label of the K-dimensional known sample label. a K-dimensional random noise of the K-dimensional known sample label is added to the K-dimensional known sample label to obtain a K-dimensional known noise sample label of the K-dimensional known sample label.

[0286] Optionally, the joint training of the first bottom model to be trained in the first participant and the top model to be trained in the second participant using the first sample feature data set and the processed known sample label set comprises: in a case where the processed known sample label set is the K-dimensional known noise sample label set, performing j-th round of joint training on the first bottom model to be trained and the top model to be trained, where j is a positive integer greater than or equal to 1, and the first bottom model and the top model obtained through 0-th round of training are the first bottom model to be trained and the top model to be trained without training, and the j-th round of joint training comprises: inputting S first sample feature data used in the j-th round of training in the first sample feature data set into the first bottom model obtained through (j-1)-th round of training to obtain S first sample forward outputs outputted through the j-th round of training, where S is a positive integer greater than or equal to 1; inputting the S first sample forward outputs outputted through the j-th round of training into the top model obtained through (j-1)-th round of training to obtain S K-dimensional sample inference results outputted through the j-th round of training; and in a case where the S K-dimensional sample inference results outputted through the j-th round of training and S K-dimensional known noise sample labels used in the j-th round of training satisfy the first end condition, ending the training to obtain the first bottom model and the top model, and the K-dimensional known noise sample labels used in the j-th round of training have the corresponding relationship with the first sample feature data used in the j-th round of training.

[0287] Optionally, the method further comprises: in a case where the S K-dimensional sample inference results outputted through the j-th round of training and the S K-dimensional known noise sample labels used in the j-th round of training do not satisfy the first end condition, adjusting model parameters of the first bottom model obtained through the (j-1)-th round of training and the top model obtained through the (j-1)-th round of training through back propagation.

[0288] Optionally, the adjusting of the model parameters of the top model obtained through the (j-1)-th round of training through back propagation comprises: obtaining, by the second participant, S first gradients through S loss functions between the S K-dimensional sample inference results and the S K-dimensional known noise sample labels and the S K-dimensional sample inference results, and adjusting the model parameters of the top model obtained through the (j-1)-th round of training through the S first gradients.

[0289] The adjustment of the first bottom model obtained in the (j-1)th training round through backpropagation includes: the second participant obtaining S third gradients through S loss functions between the inference results of the S K-dimensional samples and the labels of the S K-dimensional known noise samples, and the forward output of the S first samples; the second participant sending the S third gradients to the first participant, and the first participant adjusting the first bottom model obtained in the (j-1)th training round through the S third gradients.

[0290] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0291] According to another aspect of the embodiments of this application, a data processing apparatus for implementing the above-described data processing method is also provided. For example... Figure 9 As shown, the device includes: a first acquisition module 92, used to acquire a first forward output and a second forward output, wherein the first forward output is the output obtained by the first bottom model in the first participant processing the first feature data, and the second forward output is the output obtained by the second bottom model in the second participant processing the second feature data; a first processing module 94, used to process the first forward output and the second forward output through the top model in the second participant to obtain a K-dimensional inference result, wherein K is a positive integer greater than 2; and a first dimensionality reduction module 96, used to reduce the dimensionality of the K-dimensional inference result through the dimensionality reduction module in the second participant when K equals L. The M-dimensional inference result is obtained, wherein M is a positive integer less than K and greater than or equal to 2, and L is a preset dimension threshold parameter; wherein the first bottom model, the second bottom model, and the top model are obtained by jointly training the first bottom model to be trained in the first participant, the second bottom model to be trained in the second participant, and the top model to be trained using the first sample feature data set, the second sample feature data set, and the processed known sample label set; the processed known sample label set is the sample label set obtained by processing each M-dimensional known sample label in the M-dimensional known sample label set, and the sample label processing includes adding random noise.

[0292] With regard to the apparatus in the above embodiments, the specific manners in which various modules perform operations have been described in detail in the embodiments of the method, and thus will not be described in detail here.

[0293] According to another aspect of embodiments of the present application, a data processing apparatus is also provided, which comprises: Figure 10 As shown in the figure, the apparatus comprises: a second acquisition module 1002, configured to acquire a first sample feature data set, a second sample feature data set and an M-dimensional known sample label set having a corresponding relationship, wherein each M-dimensional known sample label in the M-dimensional known sample label set is used to represent an actual sample label; a second processing module 1004, configured to perform sample label processing on each M-dimensional known sample label in the M-dimensional known sample label set to obtain a processed known sample label set, wherein the sample label processing comprises adding random noise; and a first training module 1006, configured to perform joint training on a first bottom model to be trained in a first participant, a second bottom model to be trained in a second participant and a top model to be trained using the first sample feature data set, the second sample feature data set and the processed known sample label set until a first end condition is met, and then end the training to obtain the first bottom model, the second bottom model and the top model.

[0294] With regard to the apparatus in the above embodiments, the specific manners in which various modules perform operations have been described in detail in the embodiments of the method, and thus will not be described in detail here.

[0295] According to another aspect of embodiments of the present application, a data processing apparatus is also provided, which comprises: a third acquisition module, configured to acquire a first forward output, wherein the first forward output is an output obtained by processing first feature data by a first bottom model in a first participant; a third processing module, configured to process the first forward output by a top model in a second participant to obtain a K-dimensional inference result, wherein K is a positive integer greater than 2; and a second dimension reduction module, configured to perform dimension reduction on the K-dimensional inference result by a dimension reduction module in the second participant to obtain an M-dimensional inference result in the case that K is equal to L, wherein M is a positive integer greater than or equal to 2 and smaller than K, and L is a preset dimension threshold parameter; wherein the first bottom model and the top model are obtained by performing joint training on a first bottom model to be trained in the first participant and a top model to be trained in the second participant using a first sample feature data set and a processed known sample label set, the processed known sample label set is a sample label set obtained by performing sample label processing on each M-dimensional known sample label in an M-dimensional known sample label set, and the sample label processing comprises adding random noise.

[0296] As to the apparatus in the above-mentioned embodiments, the specific manners in which various modules perform operations have been described in detail in the embodiments related to the method, and thus will not be described here in detail.

[0297] According to another aspect of the embodiments of the present application, a data processing apparatus is further provided, which comprises: a fourth acquisition module, configured to acquire a first sample feature data set having a corresponding relationship, and an M-dimensional known sample label set, wherein each M-dimensional known sample label in the M-dimensional known sample label set is used to represent an actual sample label; a fourth processing module, configured to perform sample label processing on each M-dimensional known sample label in the M-dimensional known sample label set to obtain a processed known sample label set, wherein the sample label processing comprises: adding random noise; and a second training module, configured to perform joint training on a first bottom model to be trained in the first participant and a top model to be trained in the second participant using the first sample feature data set and the processed known sample label set until a first end condition is met, and then end the training to obtain the first bottom model and the top model.

[0298] As to the apparatus in the above-mentioned embodiments, the specific manners in which various modules perform operations have been described in detail in the embodiments related to the method, and thus will not be described here in detail.

[0299] According to an aspect of the present application, a computer program product is provided, which includes computer programs / instructions containing program codes for executing the method shown in the flow chart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication part 1109, and / or installed from the detachable medium 1111. When the computer program is executed by the central processing unit 1101, various functions provided by the embodiments of the present application are executed.

[0300] The above-mentioned serial numbers of the embodiments of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0301] Figure 11 The computer system structure block diagram of the electronic device for implementing the embodiments of the present application is schematically shown.

[0302] It should be noted that, Figure 11 The computer system 1100 of the electronic device shown is only an example, and should not bring any limitation to the functions and use range of the embodiments of the present application.

[0303] As Figure 11As shown, the computer system 1100 includes a central processing unit 1101 (CPU), which can execute various appropriate actions and processes according to programs stored in a read-only memory 1102 (ROM) or loaded from a storage section 1108 into a random access memory 1103 (RAM). Various programs and data required for system operation are also stored in the random access memory 1103. The central processing unit 1101, the read-only memory 1102, and the random access memory 1103 are connected to each other through a bus 1104. An input / output interface 1105 (I / O interface) is also connected to the bus 1104.

[0304] The following components are connected to the input / output interface 1105: an input section 1106 including a keyboard, a mouse, etc.; an output section 1107 including a display such as a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1108 including a hard disk, etc.; and a communication section 1109 including a network interface card such as a local area network card, a modem, etc. The communication section 1109 performs communication processing via a network such as the Internet. A drive 1110 is also connected to the input / output interface 1105 as necessary. A removable media 1111 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc. is mounted on the drive 1110 as necessary, so that a computer program read therefrom is installed in the storage section 1108 as necessary.

[0305] In particular, according to embodiments of the present application, the processes described in the various method flowcharts can be implemented as a computer software program. For example, embodiments of the present application include a computer program product comprising a computer program carried on a computer readable medium, the computer program containing program code for executing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network by the communication section 1109, and / or installed from the removable media 1111. When the computer program is executed by the central processing unit 1101, various functions defined in the systems of the present application are performed.

[0306] According to still another aspect of embodiments of the present application, there is also provided an electronic device for implementing the above-described data processing method, which can be Figure 1 the terminal device or a server as shown. The present embodiment is described by way of example with the electronic device being a terminal device. As Figure 12As shown, the electronic device includes a memory 1202 in which a computer program is stored, and a processor 1204 configured to execute the steps of any of the method embodiments described above by the computer program.

[0307] Optionally, in this embodiment, the electronic device described above can be located in at least one of a plurality of network devices of a computer network.

[0308] Optionally, in this embodiment, the processor described above can be configured to execute the following steps by the computer program:

[0309] S1, obtaining a first forward output and a second forward output, wherein the first forward output is an output obtained by processing first feature data by a first bottom model in a first participant, and the second forward output is an output obtained by processing second feature data by a second bottom model in a second participant;

[0310] S2, processing the first forward output and the second forward output by a top model in the second participant to obtain a K-dimensional inference result, wherein K is a positive integer greater than 2;

[0311] S3, in the case where K is equal to L, reducing the dimension of the K-dimensional inference result by a dimension reduction module in the second participant to obtain an M-dimensional inference result, wherein M is a positive integer less than K and greater than or equal to 2, and L is a preset dimension threshold parameter; wherein the first bottom model, the second bottom model and the top model are obtained by jointly training a to-be-trained first bottom model in the first participant, a to-be-trained second bottom model in the second participant and a to-be-trained top model using a first sample feature data set, a second sample feature data set and a processed known sample label set, the processed known sample label set is a sample label set obtained by performing sample label processing on each M-dimensional known sample label in an M-dimensional known sample label set, and the sample label processing includes adding random noise.

[0312] Optionally, in this embodiment, the processor described above can be further configured to execute the following steps by the computer program:

[0313] S1, obtaining a first sample feature data set, a second sample feature data set and an M-dimensional known sample label set having a corresponding relationship, wherein each M-dimensional known sample label in the M-dimensional known sample label set is used to represent an actual sample label;

[0314] S2, performing sample label processing on each M-dimensional known sample label in the set of M-dimensional known sample labels to obtain a set of processed known sample labels, wherein the sample label processing comprises adding random noise;

[0315] S3, jointly training a first bottom model in the first participant, a second bottom model in the second participant, and a top model to be trained until a first end condition is met, ending the training, to obtain the first bottom model, the second bottom model, and the top model, using the first set of sample feature data, the second set of sample feature data, and the set of processed known sample labels.

[0316] Optionally, in this embodiment, the processor can also be configured to perform the following steps by means of a computer program:

[0317] S1, obtaining a first forward output, wherein the first forward output is an output obtained by processing first feature data by a first bottom model in the first participant;

[0318] S2, processing the first forward output by a top model in the second participant to obtain a K-dimensional inference result, wherein K is a positive integer greater than 2;

[0319] S3, in the case where K is equal to L, reducing the dimension of the K-dimensional inference result by a dimension reduction module in the second participant to obtain an M-dimensional inference result, wherein M is a positive integer less than K and greater than or equal to 2, and L is a preset dimension threshold parameter; wherein the first bottom model and the top model are obtained by jointly training a first bottom model to be trained in the first participant and a top model to be trained in the second participant using a first set of sample feature data and a set of processed known sample labels, the set of processed known sample labels is obtained by performing sample label processing on each M-dimensional known sample label in a set of M-dimensional known sample labels, and the sample label processing comprises adding random noise.

[0320] Optionally, in this embodiment, the processor can also be configured to perform the following steps by means of a computer program:

[0321] S1, obtaining a first set of sample feature data having a corresponding relationship and a set of M-dimensional known sample labels, wherein each M-dimensional known sample label in the set of M-dimensional known sample labels is used to represent an actual sample label;

[0322] S2, performing sample label processing on each M-dimensional known sample label in the set of M-dimensional known sample labels to obtain a set of processed known sample labels, wherein the sample label processing comprises adding random noise;

[0323] S3, using the first sample feature data set, the processed known sample label set, jointly training the first bottom model to be trained in the first participant, the top model to be trained in the second participant, until a first end condition is met, ending the training, obtaining the first bottom model, and the top model.

[0324] Optionally, those skilled in the art can understand that, Figure 12 The structure shown is only schematic, and the electronic device can also be a smart phone (such as an Android phone, an iOS phone, etc.), a tablet computer, a palm computer, a Mobile Internet Device (MID), a PAD, or the like. Figure 12 It does not limit the structure of the electronic device. For example, the electronic device can further include more or less components (such as a network interface, etc.) than Figure 12 shown, or have a different configuration than Figure 12 shown.

[0325] The memory 1202 can be used to store software programs and modules, such as program instructions / modules corresponding to the data processing method and device in the embodiments of the present application. The processor 1204 executes various functions and data processing by running the software programs and modules stored in the memory 1202, that is, implements the above-mentioned data processing method. The memory 1202 can include a high-speed random access memory, and can also include a non-volatile memory, such as one or more magnetic storage devices, flash memories, or other non-volatile solid-state memories. In some examples, the memory 1202 can further include a memory remotely arranged with respect to the processor 1204, which can be connected to the terminal through a network. Examples of the above-mentioned network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof. The memory 1202 can specifically but not limited to be used to store first feature data, second feature data, and the like. As an example, as Figure 12 shown, the memory 1202 can include but is not limited to the first acquisition module 92, the first processing module 94, and the first dimension reduction module 96 in the above-mentioned data processing device. In addition, other module units in the above-mentioned data processing device can also be included but not limited to, which will not be described in detail in this example.

[0326] Optionally, the transmission device 1206 is configured to receive or send data via a network. Examples of the network can include a wired network and a wireless network. In an example, the transmission device 1206 includes a network interface controller (NIC) which can be connected to other network devices and routers through a network cable to communicate with the Internet or a local area network. In an example, the transmission device 1206 is a radio frequency (RF) module which is configured to communicate with the Internet in a wireless manner.

[0327] In addition, the electronic device further includes a display 1208 configured to display the M-dimensional inference result, and a connection bus 1210 configured to connect various module components in the electronic device.

[0328] In other embodiments, the terminal device or the server can be a node in a distributed system, and the distributed system can be a blockchain system formed by the plurality of nodes connected through network communication. The nodes can form a peer-to-peer (P2P) network, and any computing device, such as a server or a terminal, can become a node in the blockchain system by adding the P2P network.

[0329] According to an aspect of the present disclosure, a computer readable storage medium is provided. A processor of a computer device reads computer instructions from the computer readable storage medium. The processor executes the computer instructions to cause the computer device to perform the data processing method provided in the various optional implementation manners.

[0330] Optionally, in the embodiment, the computer readable storage medium can be configured to store a computer program for performing the following steps:

[0331] S1, obtaining a first forward output and a second forward output, wherein the first forward output is an output obtained by processing first feature data by a first bottom model in a first participant, and the second forward output is an output obtained by processing second feature data by a second bottom model in a second participant;

[0332] S2, processing the first forward output and the second forward output by a top model in the second participant to obtain a K-dimensional inference result, wherein K is a positive integer greater than 2;

[0333] S3, in the case that the K is equal to the L, performing dimension reduction on the K-dimensional inference result by a dimension reduction module in the second participant to obtain an M-dimensional inference result, where the M is less than the K and greater than or equal to 2, and the L is a preset dimension threshold parameter; wherein the first bottom model, the second bottom model and the top model are obtained by jointly training the to-be-trained first bottom model in the first participant, the to-be-trained second bottom model in the second participant and the to-be-trained top model using the first sample feature data set, the second sample feature data set and the processed known sample label set, and the processed known sample label set is obtained by performing sample label processing on each M-dimensional known sample label in the M-dimensional known sample label set, and the sample label processing includes adding random noise.

[0334] Optionally, in the embodiment, the computer readable storage medium can be configured to store a computer program for performing the following steps:

[0335] S1, obtaining a first sample feature data set, a second sample feature data set and an M-dimensional known sample label set having a corresponding relationship, wherein each M-dimensional known sample label in the M-dimensional known sample label set is used to represent an actual sample label;

[0336] S2, performing sample label processing on each M-dimensional known sample label in the M-dimensional known sample label set to obtain a processed known sample label set, wherein the sample label processing includes adding random noise;

[0337] S3, jointly training a to-be-trained first bottom model in a first participant, a to-be-trained second bottom model in a second participant and a to-be-trained top model using the first sample feature data set, the second sample feature data set and the processed known sample label set until a first end condition is met, ending the training, and obtaining the first bottom model, the second bottom model and the top model.

[0338] Optionally, in the embodiment, the computer readable storage medium can be configured to store a computer program for performing the following steps:

[0339] S1, obtaining a first forward output, wherein the first forward output is an output obtained by processing first feature data by a first bottom model in a first participant;

[0340] S2, processing the first forward output by a top model in the second participant to obtain a K-dimensional inference result, wherein the K is a positive integer greater than 2;

[0341] S3, in the case that the K is equal to the L, dimension reduction is performed on the K-dimensional inference result by a dimension reduction module in the second participant to obtain an M-dimensional inference result, where the M is less than the K and greater than or equal to 2, the L is a preset dimension threshold parameter; wherein the first bottom model and the top model are obtained by jointly training the to-be-trained first bottom model in the first participant and the to-be-trained top model in the second participant using a first sample feature data set and a processed known sample label set, the processed known sample label set is obtained by performing sample label processing on each M-dimensional known sample label in an M-dimensional known sample label set, and the sample label processing includes adding random noise.

[0342] Optionally, in the embodiment, the computer readable storage medium can be configured to store a computer program for performing the following steps:

[0343] S1, obtaining a first sample feature data set having a corresponding relationship and an M-dimensional known sample label set, where each M-dimensional known sample label in the M-dimensional known sample label set is used to represent an actual sample label;

[0344] S2, performing sample label processing on each M-dimensional known sample label in the M-dimensional known sample label set to obtain a processed known sample label set, where the sample label processing includes adding random noise;

[0345] S3, jointly training a to-be-trained first bottom model in the first participant and a to-be-trained top model in the second participant using the first sample feature data set and the processed known sample label set until a first end condition is met, ending the training, and obtaining the first bottom model and the top model.

[0346] Optionally, in the embodiment, a person of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the hardware related to the terminal device by a program, and the program can be stored in a computer readable storage medium, and the storage medium can include a flash disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0347] The sequence numbers of the above embodiments of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0348] The integrated units in the above embodiments, if implemented in the form of software function units and sold or used as independent products, can be stored in the above computer-readable storage medium. Based on such understanding, the technical solutions of the present application, essentially or in other words, the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a number of instructions to make one or more computer devices (which can be personal computers, servers or network devices, etc.) execute all or part of the steps of the methods described in the various embodiments of the present application.

[0349] In the above embodiments of the present application, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0350] In the several embodiments provided in the present application, it should be understood that the disclosed client can be implemented by other means. Among them, the above-mentioned device embodiments are only schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the shown or discussed mutual ones can be indirect coupling or communication connection through some interfaces, units or modules, and can be electrical or other forms.

[0351] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or they can be distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0352] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The above integrated unit can be realized in the form of hardware or in the form of software function unit.

[0353] The above is only the preferred embodiment of the present application, and it should be pointed out that for those skilled in the art, without departing from the principle of the present application, a number of improvements and refinements can be made, which should be regarded as the protection scope of the present application.

Claims

1. A data processing method, characterized in that, include: Obtain a first forward output and a second forward output, wherein the first forward output is the output obtained by the first bottom model in the first participant processing the first feature data, and the second forward output is the output obtained by the second bottom model in the second participant processing the second feature data. The first forward output and the second forward output are processed by the top model in the second participant to obtain a K-dimensional inference result, where K is a positive integer greater than 2; When K equals L, the K-dimensional reasoning result is reduced in dimensionality by the dimensionality reduction module in the second participant to obtain an M-dimensional reasoning result, wherein M is less than K and is a positive integer greater than or equal to 2, and L is a preset dimensionality threshold parameter. The first bottom model, the second bottom model, and the top model are obtained by jointly training the first bottom model to be trained in the first participant, the second bottom model to be trained in the second participant, and the top model to be trained using the first sample feature data set, the second sample feature data set, and the processed known sample label set. The processed known sample label set is a sample label set obtained by processing each M-dimensional known sample label in the M-dimensional known sample label set. The sample label processing includes adding random noise.

2. The method according to claim 1, characterized in that, Before obtaining the first forward output and the second forward output, the method further includes: Obtain a first sample feature data set, a second sample feature data set, and an M-dimensional known sample label set that have corresponding relationships, wherein each M-dimensional known sample label in the M-dimensional known sample label set is used to represent the actual sample label; Each M-dimensional known sample label in the M-dimensional known sample label set is subjected to sample label processing to obtain a processed known sample label set, wherein the sample label processing includes: adding random noise; The first sample feature data set, the second sample feature data set, and the processed known sample label set are used to jointly train the first bottom model to be trained in the first participant, the second bottom model to be trained in the second participant, and the top model to be trained until the first termination condition is met, and the training ends, thus obtaining the first bottom model, the second bottom model, and the top model.

3. The method according to claim 2, characterized in that, The step of processing each M-dimensional known sample label in the M-dimensional known sample label set to obtain the processed known sample label set includes: When M is less than L, the M-dimensional known sample label set is increased in dimensionality using a dimensionality-upgrading module to obtain a K-dimensional known sample label set; random noise is added to each K-dimensional known sample label in the K-dimensional known sample label set to obtain a K-dimensional known noise sample label set, wherein the processed known sample label set includes the K-dimensional known noise sample label set; or... When M is greater than or equal to L, random noise is added to each of the M-dimensional known sample labels in the M-dimensional known sample label set to obtain an M-dimensional known noise sample label set, wherein the processed known sample label set includes the M-dimensional known noise sample label set.

4. The method according to claim 3, characterized in that, The step of adding random noise to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain a K-dimensional known noise sample label set includes: adding randomly distributed noise to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set; or, adding noise of randomly opposite labels to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set; or, adding randomly opposite labels and randomly distributed noise to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set. Adding random noise to each of the M-dimensional known sample labels in the M-dimensional known sample label set to obtain an M-dimensional known noise sample label set includes: adding randomly distributed noise to each of the M-dimensional known sample labels in the M-dimensional known sample label set to obtain the M-dimensional known noise sample label set; or, adding noise of randomly opposite labels to each of the M-dimensional known sample labels in the M-dimensional known sample label set to obtain the M-dimensional known noise sample label set; or, adding random opposite labels and randomly distributed noise to each of the M-dimensional known sample labels in the M-dimensional known sample label set to obtain the M-dimensional known noise sample label set.

5. The method according to claim 4, characterized in that, Randomly distributed noise is added to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set, including: For each of the K-dimensional known sample labels in the set of K-dimensional known sample labels, the following steps are performed, where each of the K-dimensional known sample labels during the execution of the following steps is the current K-dimensional known sample label: K-dimensional random noise is generated randomly through a specified random distribution, and then processed by the softmax function to obtain K-dimensional random noise. The sum of the K-dimensional random noise and the current K-dimensional known sample label is determined as the K-dimensional known noise sample label of the current K-dimensional known sample label.

6. The method according to claim 4, characterized in that, Add random, opposing label noise to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set, including: For each of the K-dimensional known sample labels in the set of K-dimensional known sample labels, the following steps are performed, where each of the K-dimensional known sample labels during the execution of the following steps is the current K-dimensional known sample label: Obtain the opposite label of the current K-dimensional known sample label, and obtain the K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; The K-dimensional complementary sample labels randomly selected from the set of K-dimensional complementary sample labels are determined as the current K-dimensional known sample labels. 3D random noise; The sum of the K-dimensional random noise and the current K-dimensional known sample label is determined as the K-dimensional known noise sample label of the current K-dimensional known sample label.

7. The method according to claim 4, characterized in that, Add random opposing labels and randomly distributed noise to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set, including: For each of the K-dimensional known sample labels in the set of K-dimensional known sample labels, the following steps are performed, where each of the K-dimensional known sample labels during the execution of the following steps is the current K-dimensional known sample label: Random perturbations of the labels of the current K-dimensional known samples are generated randomly using a specified random distribution; Obtain the opposite label of the current K-dimensional known sample label, and obtain the K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; The product of a randomly selected K-dimensional contrastive sample label from the K-dimensional contrastive sample label set and the random perturbation of the current K-dimensional known sample label is determined as the K-dimensional random noise of the current K-dimensional known sample label. The The sum of the random noise and the current K-dimensional known sample label is determined as the K-dimensional known noise sample label of the current K-dimensional known sample label.

8. The method according to claim 3, characterized in that, The dimension-upgrading module performs dimension-upgrading on the M-dimensional known sample label set, including: obtaining K K-dimensional one-hot encoded vectors; randomly dividing the K K-dimensional one-hot encoded vectors into M groups, or dividing them into M groups according to a preset rule, to obtain M groups of K-dimensional one-hot encoded vectors; and mapping the M-dimensional known sample labels to the M groups of K-dimensional one-hot encoded vectors to obtain K-dimensional known sample labels corresponding to the M-dimensional known sample labels. The dimensionality reduction of the K-dimensional inference result is performed by the dimensionality reduction module in the second participant, including: performing an inner product operation between the K-dimensional inference result and M K-dimensional decoding vectors to obtain the M-dimensional inference result, wherein the M K-dimensional decoding vectors are decoding vectors obtained based on the M sets of K-dimensional one-hot encoding.

9. A data processing method, characterized in that, include: Obtain a first sample feature data set, a second sample feature data set, and an M-dimensional known sample label set that have a corresponding relationship, wherein each M-dimensional known sample label in the M-dimensional known sample label set is used to represent the actual sample label, and M is a positive integer greater than or equal to 2; When M is less than L, the M-dimensional known sample label set is upgraded by the dimensionality upgrade module to obtain a K-dimensional known sample label set, where L is a preset dimensionality threshold parameter and K is a positive integer greater than M; Random noise is added to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain a K-dimensional known noise sample label set. The first sample feature data set, the second sample feature data set, and the K-dimensional known noise sample label set are used to jointly train the first bottom model to be trained in the first participant, the second bottom model to be trained in the second participant, and the top model to be trained until the first termination condition is met, and the training ends, resulting in the first bottom model, the second bottom model, and the top model.

10. The method according to claim 9, characterized in that, After obtaining the first sample feature data set, the second sample feature data set, and the M-dimensional known sample label set with corresponding relationships, the process includes: When M is greater than or equal to L, random noise is added to each of the M-dimensional known sample labels in the M-dimensional known sample label set to obtain an M-dimensional known noise sample label set. Using the first sample feature data set, the second sample feature data set, and the M-dimensional known noise sample label set, the first bottom model to be trained in the first participant, the second bottom model to be trained in the second participant, and the top model to be trained are jointly trained until the first termination condition is met, and the training ends, thus obtaining the first bottom model, the second bottom model, and the top model.

11. The method according to claim 10, characterized in that, The step of adding random noise to each K-dimensional known sample label in the K-dimensional known sample label set to obtain a K-dimensional known noise sample label set includes: adding randomly distributed noise to each K-dimensional known sample label in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set; or, adding noise of randomly opposite labels to each K-dimensional known sample label in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set; or, adding random opposite labels and randomly distributed noise to each K-dimensional known sample label in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set. Adding random noise to each of the M-dimensional known sample labels in the M-dimensional known sample label set to obtain an M-dimensional known noise sample label set includes: adding randomly distributed noise to each of the M-dimensional known sample labels in the M-dimensional known sample label set to obtain the M-dimensional known noise sample label set; or, adding noise of randomly opposite labels to each of the M-dimensional known sample labels in the M-dimensional known sample label set to obtain the M-dimensional known noise sample label set; or, adding random opposite labels and randomly distributed noise to each of the M-dimensional known sample labels in the M-dimensional known sample label set to obtain the M-dimensional known noise sample label set.

12. The method according to claim 10, characterized in that, Randomly distributed noise is added to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set, including: For each of the K-dimensional known sample labels in the set of K-dimensional known sample labels, the following steps are performed, where each of the K-dimensional known sample labels during the execution of the following steps is the current K-dimensional known sample label: K-dimensional random noise is generated randomly through a specified random distribution, and then processed by the softmax function to obtain K-dimensional random noise. The sum of the K-dimensional random noise and the current K-dimensional known sample label is determined as the K-dimensional known noise sample label of the current K-dimensional known sample label.

13. The method according to claim 11, characterized in that, Add random, opposing label noise to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set, including: For each of the K-dimensional known sample labels in the set of K-dimensional known sample labels, the following steps are performed, where each of the K-dimensional known sample labels during the execution of the following steps is the current K-dimensional known sample label: Obtain the opposite label of the current K-dimensional known sample label, and obtain the K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; The K-dimensional opposing sample labels randomly selected from the K-dimensional opposing sample label set are determined as the K-dimensional random noise of the current K-dimensional known sample labels; The sum of the K-dimensional random noise and the current K-dimensional known sample label is determined as the K-dimensional known noise sample label of the current K-dimensional known sample label.

14. The method according to claim 11, characterized in that, Add random opposing labels and randomly distributed noise to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the K-dimensional known noise sample label set, including: For each of the K-dimensional known sample labels in the set of K-dimensional known sample labels, the following steps are performed, where each of the K-dimensional known sample labels during the execution of the following steps is the current K-dimensional known sample label: Random perturbations of the labels of the current K-dimensional known samples are generated randomly using a specified random distribution; Obtain the opposite label of the current K-dimensional known sample label, and obtain the K-dimensional opposite sample label set corresponding to the current K-dimensional known sample label; The product of a randomly selected K-dimensional contrastive sample label from the K-dimensional contrastive sample label set and the random perturbation of the current K-dimensional known sample label is determined as the K-dimensional random noise of the current K-dimensional known sample label. The sum of the K-dimensional random noise and the current K-dimensional known sample label is determined as the K-dimensional known noise sample label of the current K-dimensional known sample label.

15. A data processing method, characterized in that, include: Obtain the first forward output, wherein the first forward output is the output obtained by the first bottom model in the first participant processing the first feature data; The first forward output is processed by the top model in the second participant to obtain a K-dimensional inference result, where K is a positive integer greater than 2; When K equals L, the K-dimensional reasoning result is reduced in dimensionality by the dimensionality reduction module in the second participant to obtain an M-dimensional reasoning result, wherein M is less than K and is a positive integer greater than or equal to 2, and L is a preset dimensionality threshold parameter. The first bottom model and the top model are obtained by jointly training the first bottom model to be trained in the first participant and the top model to be trained in the second participant using the first sample feature data set and the processed known sample label set. The processed known sample label set is the sample label set obtained by processing each M-dimensional known sample label in the M-dimensional known sample label set. The sample label processing includes adding random noise.

16. The method according to claim 15, characterized in that, Before obtaining the first forward output, the method further includes: Obtain a first set of sample feature data with corresponding relationships and an M-dimensional known sample label set, wherein each M-dimensional known sample label in the M-dimensional known sample label set is used to represent the actual sample label; Each M-dimensional known sample label in the M-dimensional known sample label set is subjected to sample label processing to obtain a processed known sample label set, wherein the sample label processing includes: adding random noise; The first sample feature data set and the processed known sample label set are used to jointly train the first bottom model to be trained in the first participant and the top model to be trained in the second participant until the first termination condition is met, and the training ends to obtain the first bottom model and the top model.

17. A data processing method, characterized in that, include: Obtain a first set of sample feature data with corresponding relationships and an M-dimensional known sample label set, wherein each M-dimensional known sample label in the M-dimensional known sample label set is used to represent the actual sample label, and M is a positive integer greater than or equal to 2; When M is less than L, the M-dimensional known sample label set is upgraded by the dimensionality upgrade module to obtain a K-dimensional known sample label set, where L is a preset dimensionality threshold parameter and K is a positive integer greater than M; Random noise is added to each of the K-dimensional known sample labels in the K-dimensional known sample label set to obtain the processed known sample label set. The first sample feature data set and the processed known sample label set are used to jointly train the first bottom model to be trained in the first participant and the top model to be trained in the second participant until the first termination condition is met, and the training ends to obtain the first bottom model and the top model.

18. A data processing apparatus, characterized in that, include: The first acquisition module is used to acquire a first forward output and a second forward output, wherein the first forward output is the output obtained by the first bottom model in the first participant processing the first feature data, and the second forward output is the output obtained by the second bottom model in the second participant processing the second feature data. The first processing module is used to process the first forward output and the second forward output through the top model in the second participant to obtain a K-dimensional inference result, wherein K is a positive integer greater than 2; The first dimensionality reduction module is used to reduce the dimensionality of the K-dimensional inference result by the dimensionality reduction module in the second participant when K equals L, to obtain an M-dimensional inference result, wherein M is less than K and is a positive integer greater than or equal to 2, and L is a preset dimensionality threshold parameter. The first bottom model, the second bottom model, and the top model are obtained by jointly training the first bottom model to be trained in the first participant, the second bottom model to be trained in the second participant, and the top model to be trained using the first sample feature data set, the second sample feature data set, and the processed known sample label set. The processed known sample label set is a sample label set obtained by processing each M-dimensional known sample label in the M-dimensional known sample label set. The sample label processing includes adding random noise.

19. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein the program can be executed by a terminal device or computer to perform the method described in any one of claims 1 to 8, 9 to 14, 15 to 16, or 17.

20. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to perform the method described in any one of claims 1 to 8, 9 to 14, 15 to 16, or 17 via the computer program.

Images