A Federated Learning Method and System Robust to Mixed Noise

Abstract

This invention proposes a federated learning method robust to mixed noise, comprising: sending local metric model parameters updated by the client based on local training data and a subjective logistic loss function to the server; obtaining global metric model parameters calculated by the server based on the local metric model parameters updated by the client and other clients, and the corresponding sample data volume; the client calculating the subjective logistic loss and local Gaussian mixture distribution for each training sample based on the global metric model parameters and local training data, and sending them to the server; obtaining interval thresholds calculated by the server based on multiple local Gaussian mixture distributions, and performing mixed noise identification on the training data based on the interval thresholds and local Gaussian mixture distributions, thereby filtering and correcting the identified open set noise and closed set noise respectively. This invention also proposes a federated learning system robust to mixed noise, and a data processing device for federated learning.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and more specifically to a federated learning method and system that is robust to mixed noise. Background Technology

[0002] With the development of distributed machine learning and big data analytics, federated learning has emerged as a novel distributed machine learning framework, supporting multiple clients (institutions) to collaboratively train models while ensuring data privacy. During model training, the server and clients only exchange intermediate parameters; each client does not need to upload any raw data. In practical federated learning scenarios, while the addition of multiple clients brings more knowledge, it also increases the risk of label noise, impacting model performance. Because federated learning involves "data available but not visible," federated learning with noise faces greater challenges. In real-world scenarios, label noise is mainly divided into two categories: open-set noise and closed-set noise. For example, in an AI-assisted pneumonia diagnosis task, a chest CT scan labeled as pneumonia is an open-set noise sample, its true category not belonging to this classification task; while a normal lung CT scan labeled as pneumonia is a closed-set noise sample, its true category (normal) is included in this classification task. Overall, open-set noise is suitable for noise filtering methods, while closed-set noise, due to its reusability, is suitable for label inference methods, reducing information loss. Due to the subjectivity of annotation and annotation bias, the noise level and noise components of each client are heterogeneous. Therefore, in federated learning, it is crucial to identify the noise type for different clients and then process them separately.

[0003] Existing techniques for addressing label noise often rely on an assumption about the type of noise. For open-set noise, since the original features are irrelevant to the classification task, methods such as sample selection and noise weighting are commonly used for noise filtering. For closed-set noise, since the original features belong to the classification task and have high reusability, methods such as label inference and loss function correction are commonly used for noise correction. However, real-world federated learning scenarios are more complex. The heterogeneity of client sources means that label noise is very likely to appear in the form of a mixture of open and closed sets, and the noise level varies from client to client. Existing methods cannot provide a robust federated learning solution for mixed noise. How to identify and process mixed noise separately in federated learning scenarios is a critical problem that urgently needs to be solved. Summary of the Invention

[0004] To address the above problems, this invention proposes a federated learning method robust to mixed noise, comprising:

[0005] Step 1: Build a local metric model on the client side, and train the local metric model for a specified number of rounds using local training data to complete the training of the local metric model. Model parameters The local metric model parameters are sent to the server; the model parameters are... The first sample size of the local training data is sent to the server; this local metric model Using the subjective logistic function as the loss function, the loss function for a given local training data i is:

[0006]

[0007] in, Dirichlet distribution The coefficients in , For the activation function ReLU with respect to the feature data of the local metric model in i The output above indicates support for the first... Evidence of class labels , represents the potential of the Dirichlet distribution, n represents the client number, and the subscript b represents the metric model;

[0008] Step 2, on the server side, use all the model parameters Step 3: On the client side, using the global metric model parameters and the local training data, obtain the subjective logistic loss value for each local training data point, thereby fitting the local Gaussian mixture distribution of all the local training data. Step 4: On the server side, using all the local Gaussian mixture distributions, generate a local Gaussian mixture discrete distribution and aggregate them to obtain a global Gaussian mixture distribution. Obtain the KS distance between the local Gaussian mixture discrete distribution and the global Gaussian mixture distribution, and obtain the discrimination threshold based on the KS distance. Step 5: On the client side, select training samples from the local training data based on the discrimination threshold. Step 6: Train the local classification model using the training dataset; correct the labels of the local training data using the local classification model, and send the local classification model parameters and the second sample size of the corrected local training data to the server; Step 7: Obtain the global classification model parameters on the server using all the local classification model parameters and the second sample size; Step 8: Update the local classification model to the global classification model on the client using the global classification model parameters; Repeat steps 6 to 8 until the global classification model of the client converges, and use the global classification model at this time as the final classification model of the client.

[0009] In the federated learning method of this invention, step 2 obtains the global metric model parameters using the following formula. :

[0010]

[0011] This indicates the amount of local training data. This represents the sum of the first sample sizes for all clients. Indicates the client sequence number. Indicates the number of clients.

[0012] In the federated learning method of this invention, step 4 includes a first discrimination threshold λ. I Second distinction threshold λ II , λ I Let λ be the inflection point value of the KS distance, used to divide the training dataset into a clean subset of data and an open-set noisy subset of data. II The minimum KS distance is used to divide the training dataset into open-set noisy data subsets and closed-set noisy data subsets.

[0013]

[0014]

[0015] For discrete global Gaussian mixture distribution The cumulative distribution function, For discrete local Gaussian mixture distribution The conditional probability distribution, .

[0016] This invention also proposes a federated learning system robust to mixed noise, comprising: a model building module, set on the client, for building a local metric model, and training the local metric model for a specified number of rounds using local training data to complete the training of the local metric model. Model parameters These are the local metric model parameters sent to the server; these model parameters... The first sample size of the local training data is sent to the server; this local metric model Using the subjective logistic function as the loss function, the loss function for a given local training data i is:

[0017]

[0018] in, Dirichlet distribution The coefficients in , For the activation function ReLU with respect to the feature data of the local metric model in i The output above indicates support for the first... Evidence of class labels , represents the potential of the Dirichlet distribution, n represents the client number, and the subscript b represents the metric model;

[0019] The metric model aggregation module, set up on the server side, is used to aggregate all parameters of the model. The system uses the first sample size to obtain global metric model parameters; a local Gaussian mixture distribution generation module, located on the client, is used to obtain the subjective logistic loss value of each local training data point using the global metric model parameters and the local training data, thereby fitting a local Gaussian mixture distribution for all the local training data; a threshold calculation module, located on the server, is used to generate a local Gaussian mixture discrete distribution using all the local Gaussian mixture distributions, aggregate them to obtain a global Gaussian mixture distribution, obtain the KS distance between the local Gaussian mixture discrete distribution and the global Gaussian mixture distribution, and obtain a discrimination threshold based on the KS distance; a noise identification module, located on the client, is used to select a training dataset from the local training data based on the discrimination threshold; and a noise correction module, located on the client. The system is used to train the local classification model using the training dataset; to correct the labels of the local training data using the local classification model; and to send the local classification model parameters and the second sample size of the corrected local training data to the server. A classification model aggregation module, located on the server, is used to obtain the global classification model parameters using all the local classification model parameters and the second sample size. A model update module, located on the client, is used to update the local classification model to a global classification model using the global classification model parameters. An iterative convergence module, located on the client, is used to repeatedly call the noise correction module, the classification model aggregation module, and the model update module sequentially until the client's global classification model converges, and uses this global classification model as the client's final classification model.

[0020] In the federated learning system of this invention, the global metric model parameters are obtained in the metric model aggregation module using the following formula. :

[0021]

[0022] This indicates the amount of local training data. This represents the sum of the first sample sizes for all clients. Indicates the client sequence number. Indicates the number of clients.

[0023] The federated learning system of the present invention, wherein the threshold calculation module includes: a first threshold calculation module, used to use the inflection point value of the KS distance as a first discrimination threshold λ. I , λ I This is used to divide the training dataset into a clean subset of data and an open-set noisy subset of data. The second threshold calculation module is used to set the minimum value of the KS distance as the second distinguishing threshold λ. II , λ II This is used to divide the training dataset into open-set noisy subsets and closed-set noisy subsets. ; For discrete global Gaussian mixture distribution The cumulative distribution function, For discrete local Gaussian mixture distribution The conditional probability distribution, .

[0024] The present invention also proposes a computer-readable storage medium storing computer-executable instructions, characterized in that, when the computer-executable instructions are executed, the federated learning method robust to mixed noise as described above is implemented.

[0025] The present invention also proposes a data processing apparatus, including a computer-readable storage medium and a processor as described above; when the data processing apparatus acts as a client or a server, the processor retrieves and executes computer-executable instructions in the computer-readable storage medium to perform federated learning robust to mixed noise. Attached Figure Description

[0026] Figure 1 This is a flowchart of a mixed-noise robust federated learning method according to an embodiment of the present invention.

[0027] Figure 2 This is a schematic diagram of a federated learning system according to an embodiment of the present invention.

[0028] Figure 3 This is a schematic diagram of a mixed-noise robust federated learning system according to an embodiment of the present invention.

[0029] Figure 4 , Figure 5 , Figure 6 This is a schematic diagram of the experimental results according to an embodiment of the present invention.

[0030] Figure 7 This is a schematic diagram of the data processing device of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0032] While researching mixed noise in federated learning, the inventors discovered that the difficulty in distinguishing mixed noise in existing technologies stems from the failure to mine the confidence intervals between open-set and closed-set noise. Noise differentiation often requires learning separability features and solving the confidence interval problem based on these features. In real-world applications of federated learning, due to labeling and sampling costs, the server struggles to provide an additional reference distribution. Some methods, even those that construct separability features, can only estimate confidence intervals by pre-setting fixed hyperparameters due to the lack of a reference distribution, resulting in a loss of recognition accuracy. The limitation of reference distribution in existing technologies essentially stems from neglecting the inherent client (local)-server (global) collaborative nature of federated learning. Existing technologies remain in a reference distribution-dependent mode, similar to traditional centralized methods, failing to shift the focus to the "federated" advantages of federated learning, thus necessitating reliance on an additional reference distribution.

[0033] Therefore, there is an urgent need for a federated learning system that can both identify mixed noise and not rely on additional reference distributions.

[0034] This invention provides a federated learning method robust to mixed noise, comprising: sending local metric model parameters updated by the client based on local training data and a subjective logistic loss function, along with the corresponding sample data volume, to the server; obtaining global metric model parameters calculated by the server based on the local metric model parameters updated by the client and other clients, and the corresponding sample data volume; the client calculating the subjective logistic loss for each training sample based on the global metric model parameters and local training data, fitting a local Gaussian mixture distribution, and sending the local Gaussian mixture distribution to the server; and the server calculating a global Gaussian mixture distribution based on the multiple local Gaussian mixture distributions, and then calculating the global mixture distribution based on the global sum of the local Gaussian mixture distributions. The distance calculation of the local Gaussian mixture distribution indicates the interval threshold for different clients. The client obtains the interval threshold and uses it to identify mixed noise in the training data, filtering out the identified open-set noise. The client then performs semi-supervised training on the identified closed-set noise and clean data to obtain a local classification model. Based on this local classification model, the client corrects the noise labels on the local training data and sends the local classification model to the server. The server obtains the global classification model parameters calculated based on the updated local classification models from the client and other clients, along with the corresponding sample data volume. The obtained global classification model is then used to replace the local classification model parameters. This mixed noise processing method is applied to a federated learning system, achieving mixed noise identification and correction independent of the reference distribution, effectively improving the accuracy of the federated learning classification model.

[0035] Before describing the embodiments of the present invention in detail, some of the terms used therein are explained as follows:

[0036] The client, also known as the user terminal, is the node that provides services to the client. Clients can be workstations belonging to different organizations (such as medical institutions, universities, research institutes, etc.). These workstations generally have barriers to data communication, preventing them from directly exchanging raw image data or uploading image data to a central node. Clients can also be large-scale mobile devices or other IoT devices, with raw image data stored locally. Although the mixed-noise robust federated learning method of this invention achieves excellent results in image classification, the client of this invention is not limited to any application scenario.

[0037] The server, also known as the central node, is the node that provides services to the clients. The server coordinates joint modeling among multiple clients without transmitting their original image data. For example, it supports the exchange of intermediate parameters (such as the model parameters in this invention), undertakes tasks such as parameter aggregation and distribution, and remains invisible to the clients' original image data. Similarly, the server in this invention is not limited to any specific application scenario.

[0038] A machine learning model is an artificial neural network formed by the interconnectedness of multiple neurons, simulated by processing units. It reflects the fundamental characteristics of human brain function and is a highly complex nonlinear learning system. The model updated on the client side is called the local model, while the model updated on the server side is called the global model.

[0039] To address the mixed noise identification problem and the reference distribution dependency problem mentioned above, this invention proposes a mixed noise robust federated learning system.

[0040] The objective of this invention is achieved through the following technical solution:

[0041] According to a first aspect of the present invention, a hybrid noise robust federated learning method is provided, comprising: sending local metric model parameters updated by a client based on local training image data and a subjective logistic loss function, along with the corresponding sample data volume, to a server; obtaining global metric model parameters calculated by the server based on the local metric model parameters updated by the client and other clients, and the corresponding sample data volume; having the client calculate the subjective logistic loss for each training image data sample based on the global metric model parameters and local image training data, and fitting a local Gaussian mixture distribution, and sending the local Gaussian mixture distribution to the server; and having the server calculate a global Gaussian mixture distribution based on the plurality of local Gaussian mixture distributions. The system calculates interval thresholds for different clients based on the distance between the global and local Gaussian mixture distributions. The client then uses these interval thresholds and the local Gaussian mixture distribution to identify mixed noise in the training image data. The identified open-set noise is filtered, and the identified closed-set noise is semi-supervised with clean data to obtain a local image classification model. Noise labels are then corrected on the local image training data based on the local image classification model. Finally, the server obtains the global image classification model parameters calculated by the server based on the local image classification model updated by the client and other clients, along with the corresponding sample data volume. The obtained global image classification model is then used to replace the local image classification model parameters.

[0042] In some embodiments of the present invention, the image sample data volume includes the first sample volume of local image training data used in the current update of the local metric model, and the global metric model parameters are obtained by the server by weighted summation of the local model parameters of all clients, wherein the first sample volume of the client divided by the sum of the first sample volumes of the client and other clients is used as the weight of the local metric model parameters of the client.

[0043] In some embodiments of the present invention, the step of calculating the subjective logical loss of each training sample and fitting a local Gaussian mixture distribution by the client based on the global metric model parameters and local training data includes: the client predicting local image data using the latest global metric model to obtain the subjective logical loss value of each sample in the local image training data; and fitting a local Gaussian mixture distribution based on all the subjective logical loss values ​​generated by the local image training data.

[0044] In some embodiments of the present invention, the global Gaussian mixture distribution is obtained by the server sampling and aggregating the local Gaussian mixture distributions of all clients. The local Gaussian mixture distribution of each client is sampled to generate a discrete distribution. The discrete distribution is then combined with the discrete distributions of other clients to obtain a global discrete distribution. The global Gaussian mixture distribution is then fitted based on the global discrete distribution.

[0045] In some embodiments of the present invention, the step of calculating the interval thresholds for different clients based on the distance calculation of the global and local Gaussian mixture distributions includes: generating a conditional distribution controlled by a λ parameter based on the local Gaussian mixture distribution, continuously increasing λ, calculating the KS distance between the conditional distribution and the global Gaussian mixture distribution, wherein threshold I is equal to the λ value corresponding to the inflection point value during the change of the KS distance, and threshold II is equal to the λ value corresponding to the minimum value during the change of the KS distance. Since each client uploads a different local Gaussian mixture distribution, different clients will obtain different threshold combinations, i.e., different thresholds I and threshold II.

[0046] In some embodiments of the present invention, the step of performing mixed noise recognition on the local image training data of the client based on the interval threshold and the local Gaussian mixture distribution includes: predicting the subjective logical loss value generated from the local image training data according to the local Gaussian mixture distribution to obtain the prediction result; calculating the probability that each sample belongs to clean sample, open set noise, and closed set noise according to the prediction result and the threshold combination; predicting the type of the sample according to the largest probability value among the three to obtain the clean sample set, the open set noise sample set, and the closed set noise sample set.

[0047] In some embodiments of the present invention, the sample data volume includes a second sample volume consisting of the sum of the clean sample volume and the closed-set noise sample volume in the local training data used in this update of the local classification model, wherein the second sample volume of the client divided by the sum of the second sample volumes of the client and other clients is used as the weight of the local classification model parameters of the client.

[0048] In some embodiments of the present invention, the step of semi-supervised training of the identified closed-set noise and clean data to obtain a local image classification model includes: constructing an enhanced supervised sample set. and enhanced unsupervised sample sets In this process, supervised samples augment clean data in batches, with data augmented from the same sample sharing the same supervision signal. This supervision signal equals the sample's clean probability multiplied by the original label's one-hot encoding, plus 1 - the clean probability multiplied by the local classification model's prediction of the sample. Similarly, unsupervised samples augment closed-set noise in batches, with data augmented from the same sample sharing the same pseudo-supervision signal. This pseudo-supervision signal equals the local classification model's average prediction of the augmented data. The augmented supervised samples and augmented unsupervised samples (including the pseudo-supervision signal) are concatenated into a new dataset, which is then randomly shuffled to obtain a shuffled dataset. Finally, the augmented supervised sample set and the shuffled dataset are fused to obtain a fused supervised sample set. That is, take the first few bits of the shuffled dataset. Each data point and supervised sample are matched one-to-one and weighted to obtain a fused supervised sample. The amount of data equal to the supervised sample data is used. During fusion, a parameter η is generated according to the Beta distribution. The larger value between η and 1−η is used as the weight of the supervised sample data, and the smaller value is used as the weight of the corresponding sample in the shuffled dataset. The enhanced unsupervised sample set and the shuffled dataset are fused to obtain the fused unsupervised sample set. That is, take the shuffled dataset. Each set of data points and unsupervised samples are matched one-to-one and weighted to obtain a fused unsupervised sample. The amount of data equal to the unsupervised sample set is used to generate corresponding fused unsupervised samples during fusion, following the aforementioned Beta distribution. A local image classification model is trained based on the fused supervised sample set and the fused unsupervised sample set, using the MixMatch loss function for training the local classification model.

[0049]

[0050] in, This represents the local MixMatch loss function for the nth client. This represents the loss function applied to the fused supervised sample set. This represents the loss function applied to the fusion of unsupervised sample sets. Indicates a monitoring signal The prediction results of the local image classification model on the samples Cross-entropy between Indicating a false supervisory signal The prediction results of the local image classification model on the samples The square of the 2-norm between, It is an inherent parameter in MixMatch, used to weight and fuse unsupervised sample sets as a loss function.

[0051] In some embodiments of the present invention, the step of correcting noise labels on local training data based on a local image classification model includes: performing label inference on the local training data according to the latest local image classification model trained by MixMatch and the closed-set noise probability, i.e., calculating pseudo-labels. First, an intermediate result is generated, which is equal to the prediction probability of the local classification model for the training sample multiplied by the closed-set noise probability plus the initial label one-hot encoding of the training sample multiplied by 1 - the closed-set noise probability. The category corresponding to the largest output value in the intermediate result is taken as the category indicated by the pseudo-label, and the corresponding pseudo-label one-hot encoding is generated. Each client replaces the local classification model parameters with the obtained global classification model parameters, continues to train the local classification model using the local training data after noise label correction, and updates the local classification model parameters; on the server side, the global model parameters are updated according to the local model parameters updated by multiple clients.

[0052] According to a second aspect of the present invention, a federated learning system is provided, comprising: a server and multiple clients; the server being configured to organize the multiple clients to exchange parameters in a privacy-preserving manner to generate intermediate parameters, the intermediate parameters including global metric model parameters, threshold combination parameters, and global classification model parameters; the multiple clients being configured to perform federated learning based on the intermediate parameters, and to filter and correct label noise in local training data using the method of the first aspect; and each client training a local image classification model using the local training data after filtering and correcting label noise, and updating the local image classification model parameters; the server being further configured to update the global image classification model parameters according to the local image classification model parameters updated by the multiple clients.

[0053] In some embodiments of the present invention, the client includes: a metric model training module, used to send the local metric model parameters updated by the current client based on local image training data and the corresponding sample data volume to the server; a local Gaussian mixture distribution calculation module, used to obtain the global metric model parameters updated by the server based on multiple clients, calculate the subjective logical loss value of each training sample based on the latest global metric model parameters and local image training data, and fit a local Gaussian mixture distribution to send to the server; and a noise recognition module, used to obtain the threshold combination indicating different clients calculated by the server based on the multiple local Gaussian mixture distributions, and pre-calculate the threshold combination and the local Gaussian mixture distribution. The system measures open-set noise and closed-set noise in the local image training data, filters out the open-set noise, and performs semi-supervised training on the closed-set noise sample set and the clean data sample set. The classification model training module sends the updated local image classification model parameters and corresponding sample data volume from the current client to the server. The noise correction module calculates pseudo-labels based on the prediction results of the local image classification model parameters and replaces the original labels in the local image training data with these pseudo-labels. The classification model training module also trains the current client's local image classification model using the local training data with corrected label noise and updates the local image classification model parameters.

[0054] In some embodiments of the present invention, the server includes: a metric model aggregation module, configured to acquire local metric model parameters updated based on the client's own local image training data and the corresponding sample data volume sent by multiple clients, calculate global metric model parameters based on the local metric model parameters updated by multiple clients and the corresponding sample data volume, and send them to multiple clients; a threshold calculation module, configured to acquire local Gaussian mixture distributions sent by the multiple clients, calculate a global Gaussian mixture distribution based on the multiple local Gaussian mixture distributions, and calculate corresponding threshold combinations (threshold I and threshold II) based on the KS distance between each local Gaussian mixture distribution and the global Gaussian mixture distribution; and a classification model aggregation module, configured to acquire local image classification model parameters updated based on client-filtered open set noise local image training data sent by multiple clients and the corresponding sample data volume, calculate global classification model parameters based on the local image classification model parameters updated by multiple clients and the corresponding sample data volume, and send them to multiple clients; wherein, the classification model aggregation module is further configured to acquire local image classification model parameters updated by multiple clients after correcting label noise based on the latest local image classification model parameters.

[0055] As mentioned in the background section, when conducting research on mixed noise in federated learning, the inventors discovered that the difficulty in distinguishing noise types in existing technologies stemmed from the lack of confidence intervals for open-set and closed-set noise. Distinguishing noise often requires learning separability features and solving the confidence interval problem based on these features. Since the server rarely provides an additional reference distribution, some methods, even after constructing separability features, can only estimate confidence intervals by pre-setting fixed hyperparameters, resulting in a loss of recognition accuracy. Therefore, before the model processes the data separately, a global reference distribution needs to be constructed to support comparison between each local distribution and this global distribution, thereby capturing adaptive thresholds to divide the confidence intervals.

[0056] Based on the above research, according to an embodiment of the present invention, such as Figure 1 As shown, a federated learning method based on hybrid noise robustness is provided, including execution steps S1~S12, each step is described in detail below.

[0057] Step S1: Send the local metric model parameters updated by the client based on the local training data and the corresponding sample data volume to the server.

[0058] According to one embodiment of the present invention, local training data is fed into the client's local model in batches for multiple rounds of training until the local model reaches a specified number of rounds. Training then stops, and the local model parameters are measured and sent to the server. For example, the number of iteration rounds for the local model can be specified as 10 rounds, but other implementers can set empirical values ​​based on different models. The model is iteratively updated in batches of data, with each batch containing 32 training samples. The local model parameters are measured using a subjective logistic function as the loss function. The formula for calculating the subjective logistic loss for a training sample is:

[0059] (1)

[0060] parameter Corresponding to a Dirichlet distribution The coefficients in ,in The activation function ReLU is used with respect to the local metric model in the feature data. The output above indicates support for the first... Evidence of class; This represents the potential of the Dirichlet distribution. The Dirichlet distribution is often used to describe multi-class classification problems, when there is no evidence to support a particular class. The Dirichlet distribution becomes a uniform distribution. .

[0061] The model parameters of the local metric model are updated using stochastic gradient descent, as follows:

[0062] (2)

[0063] in, This indicates the batch number of each small batch of data input into the model. This represents the parameters of the local model from the previous batch. This represents the model's loss on this batch of data. This indicates the gradient that needs to be lowered in this batch. This represents the learning rate. As the step size for gradient descent.

[0064] Step S2: Obtain the local metric model parameters updated by the server based on the client and other clients, and the global metric model parameters calculated by the corresponding first sample size.

[0065] According to an embodiment of the present invention, the global metric model parameters can be calculated according to the following formula:

[0066] (3)

[0067] in, For the latest local metric model parameters, This indicates the amount of local training data. This represents the sum of the local training data volume of all clients. Indicates the client's sequence number. This indicates the number of clients, and the subscript b indicates that the current model is a metric model.

[0068] Step S3: The client calculates the subjective logical loss value of each training sample based on the global metric model parameters and local training data, using the same calculation method as formula (1).

[0069] Step S4: The client uses a Gaussian mixture distribution to fit all subjective logistic loss values ​​to obtain a local Gaussian mixture distribution, i.e. The local Gaussian mixture distribution is then sent to the server.

[0070] Step S5: The server generates discrete local Gaussian mixture distributions based on the local Gaussian mixture distributions of all clients, and aggregates and calculates discrete global Gaussian mixture distributions based on the discrete local Gaussian mixture distributions.

[0071] According to one embodiment of the present invention, the discrete local Gaussian mixture distribution can be calculated according to the following formula:

[0072] (4)

[0073] in, Indicates a local Gaussian mixture distribution Discrete sampling points generated by sampling, co-sampled Second-rate.

[0074] According to one embodiment of the present invention, the discrete global Gaussian mixture distribution can be calculated according to the following formula:

[0075] (5)

[0076] in, Represents the global Gaussian mixture distribution fitted from the union of discrete local Gaussian mixture distributions. Discrete sampling points generated by sampling, co-sampled .

[0077] Step S6: The server determines the discrete local Gaussian mixture distribution. With discrete global Gaussian mixture distribution Calculate the KS divergence to measure the distance between the two:

[0078] (6)

[0079] in, Global loss distribution The cumulative distribution function, i.e. For local loss distribution The cumulative distribution function (CDF), i.e. . for conditional probability distribution It obviously satisfies and .

[0080] Step S7: The server determines the method based on... Conditional probability distribution and discrete global Gaussian mixture distribution The KS distance is calculated, and two distinction thresholds are calculated. First, :

[0081] (7)

[0082] in, The inflection point value corresponding to the KS distance is used to guide the client in separating clean data from open set noise; and :

[0083] (8)

[0084] in, The minimum value corresponding to the KS distance is used to guide the client in distinguishing between open set noise and closed set noise.

[0085] Step S8: Obtain the threshold combination calculated by the server based on the discrete local Gaussian mixture distribution and the discrete global Gaussian mixture distribution. and The client divides the training data of each participant into three subsets based on the threshold combination: a clean data set, an open-set noise set, and a closed-set noise set. Specifically, it utilizes a local Gaussian mixture distribution. Predict the probability of a training sample being of each type, with the clean probability being:

[0086] (9)

[0087] The probability of open set noise is:

[0088] (10)

[0089] The probability of closed-set noise is:

[0090] (11)

[0091] in, Indicates the local Gaussian mixture distribution prediction sample loss value In the The probabilities are obtained from a Gaussian distribution. Based on the obtained probabilities, a clean sample set is defined. Open set noise sample set and closed set of noise samples And filter out the open set of noisy samples.

[0092] According to one embodiment of the present invention, the formulas for calculating the clean sample set, the open-set noisy sample set, and the closed-set noisy sample set are as follows:

[0093] (12)

[0094] Step S9: The client uses MixMatch semi-supervised learning to train a local classification model on clean sample sets and closed noisy sample sets. The subscript c indicates that the current model is a classification model.

[0095] According to one embodiment of the present invention, firstly, an enhanced supervised sample set is constructed. and enhanced unsupervised sample sets In this process, supervised samples perform data augmentation on the identified clean data in batches. Data augmented from the same sample share the same supervision signal, which is equal to the clean probability of the sample multiplied by the one-hot encoding of the original label, plus 1 - the clean probability multiplied by the prediction result of the local classification model for the sample.

[0096] According to one embodiment of the present invention, unsupervised samples are also used to augment the identified closed-set noise in batches. The data augmented from the same sample share the same pseudo-supervisory signal, which is equal to the average prediction result of the local classification model for the augmented data.

[0097] According to one embodiment of the present invention, enhanced supervised samples and enhanced unsupervised samples (including pseudo-supervised signals) are concatenated into a new dataset, and the dataset is randomly shuffled to obtain a shuffled dataset.

[0098] According to one embodiment of the present invention, a fused supervised sample set is obtained by fusing the enhanced supervised sample set and the shuffled dataset. That is, take the first few bits of the shuffled dataset. Each data point and supervised sample are matched one-to-one and weighted to obtain a fused supervised sample. The amount of data is equal to the amount of supervised samples. During fusion, the parameter η is generated according to the Beta distribution. The larger value between η and 1−η is used as the weight of the supervised samples, and the smaller value is used as the weight of the corresponding samples in the shuffled dataset.

[0099] According to one embodiment of the present invention, a fused unsupervised sample set is obtained by fusing the enhanced unsupervised sample set and the shuffled dataset. That is, take the shuffled dataset. Each set of data points and unsupervised samples are matched one-to-one and weighted to obtain a fused unsupervised sample. The amount of data is equal to the amount of unsupervised samples. During fusion, corresponding fused unsupervised samples are generated according to the Beta distribution mentioned above.

[0100] According to one embodiment of the present invention, a local classification model is trained based on the fused supervised sample set and the fused unsupervised sample set, wherein the MixMatch loss function is selected when training the local classification model.

[0101] (13)

[0102] in, This represents the local MixMatch loss function for the nth client. This represents the loss function applied to the fused supervised sample set. This represents the loss function applied to the fusion of unsupervised sample sets. Indicates a monitoring signal The prediction results of the local classification model on the samples Cross-entropy between Indicating a false supervisory signal The prediction results of the local classification model on the samples The square of the 2-norm between, It is an inherent parameter in MixMatch, used to weight and fuse unsupervised sample sets as a loss function.

[0103] According to one embodiment of the present invention, the client trains a local classification model using the MixMatch loss function:

[0104] (14)

[0105] Step S10: The client performs label correction on the local training data based on the local classification model.

[0106] According to one embodiment of the present invention, the formula for calculating the pseudo-tag is:

[0107]

[0108] Among them, closed-set noise probability As a model prediction result The greater the weight of the closed-set noise probability, the more confident we are in the model's prediction results.

[0109] According to one embodiment of the present invention, the client performs one-hot encoding on the pseudo-tag:

[0110] (15)

[0111] According to one embodiment of the present invention, for corrected samples, i.e., training samples whose pseudo-labels and original labels are inconsistent, they are assigned to a clean sample set; corrected samples are removed from the closed-set noisy sample set.

[0112]

[0113] Step S11: Obtain the local classification model parameters updated by the server based on the client and other clients, and the global classification model parameters calculated by the corresponding second sample size, and replace the local classification model with the global classification model.

[0114] According to an embodiment of the present invention, the global classification model parameters can be calculated according to the following formula:

[0115] (16)

[0116] The second sample size of the client is divided by the sum of the second sample sizes of the client and other clients, which serves as the weight of the local classification model for the client.

[0117] Step S12: Repeat steps S9 to S11 until the global classification model converges.

[0118] In one embodiment, such as Figure 3 As shown, it includes at least two clients and a server.

[0119] Each client may include: a metric model training module, a local Gaussian mixture distribution calculation module and a noise identification module, a classification model training module, and a noise correction module.

[0120] The metric model training module includes a metric model training unit, which is used to send the local metric model parameters updated by the current client based on local training data and subjective logical loss function, along with the corresponding sample data volume, to the server.

[0121] The local Gaussian mixture distribution generation module may include a subjective logic loss value calculation unit and a local Gaussian mixture distribution calculation unit.

[0122] The subjective logical loss calculation unit is used to calculate the subjective logical loss value of each local sample based on the global metric model.

[0123] A local Gaussian mixture distribution generation unit is used to fit a Gaussian mixture distribution based on the subjective logical loss value.

[0124] The noise identification module includes a sample set partitioning unit and an open set noise filtering unit.

[0125] The sample set partitioning unit is used to partition a clean sample set, an open set noise sample set, and a closed set noise sample set according to the adaptive threshold and the local Gaussian mixture distribution.

[0126] The open-set noise filtering unit is used to reserve the open-set noise sample set.

[0127] The classification model training module includes a classification model training unit, which sends the local classification model parameters updated by the current client based on the local clean sample set and the closed noisy sample set, along with the corresponding amount of sample data, to the server.

[0128] The noise correction module includes a noise correction unit, which calculates pseudo-labels based on the prediction results of the local classification model parameters, replaces the original labels of the local training data with pseudo-labels, and forms a new clean sample set and a closed noisy sample set.

[0129] According to one embodiment of the present invention, the server may include: a metric model aggregation module, a threshold calculation module, and a classification model aggregation module.

[0130] The metric model aggregation module can be used to obtain local metric model parameters and corresponding sample data volumes updated based on the client's own first sample size sent by multiple clients, calculate global metric model parameters based on the local metric model parameters updated by multiple clients and the corresponding sample data volumes, and send them to multiple clients.

[0131] The threshold calculation module can be used to obtain multiple local Gaussian mixture distributions sent by multiple clients, generate a global Gaussian mixture distribution based on the multiple local Gaussian mixture distributions, calculate the adaptive threshold for different clients based on the global Gaussian mixture distribution and each local Gaussian mixture distribution, and send the corresponding threshold for each client to that client.

[0132] The classification model aggregation module can be used to obtain local classification model parameters updated by one or more clients based on the latest global classification model parameters, the latest clean sample set, and the closed-set noisy sample set, and then aggregate the updated local model parameters to obtain the global classification model parameters and send them to multiple clients.

[0133] It should be understood that although the mixed noise robust federated learning method of the present invention can achieve excellent results in image classification, it is not limited to other application scenarios, such as human body parameter (face, fingerprint, retina, iris, palm print, voice, etc.) recognition, information collection, etc., and the present invention is not limited thereto.

[0134] Figure 4 , Figure 5 , Figure 6 This is a schematic diagram of experimental results according to an embodiment of the present invention. Wherein:

[0135] Figure 4 The results shown are experimental results on the publicly available object recognition dataset CIFAR-10 under noise intensity (data noise ratio 0.4, of which open set noise ratio 0.3) according to an embodiment of the present invention.

[0136] Figure 5 The results shown are experimental results on the publicly available object recognition dataset CIFAR-10 under noise intensity (data noise ratio 0.4, of which open set noise ratio 0.5) according to an embodiment of the present invention.

[0137] Figure 6 The results shown are experimental results on the publicly available object recognition dataset CIFAR-10 under noise intensity (data noise ratio 0.4, of which open set noise ratio 0.7) according to an embodiment of the present invention.

[0138] Figure 7 This is a schematic diagram of the data processing apparatus of the present invention. Figure 7As shown, embodiments of the present invention also provide a computer-readable storage medium and a data processing apparatus. The data processing apparatus of the present invention can function as a client or a server. The computer-readable storage medium of the present invention is disposed within the data processing apparatus and stores computer-executable instructions. When the computer-executable instructions are executed by the processor of the data processing apparatus, the aforementioned federated learning method robust to mixed noise is implemented. Those skilled in the art will understand that all or part of the steps in the above method can be implemented by a program instructing related hardware (e.g., processor, FPGA, ASIC, etc.), and the program can be stored in a readable storage medium, such as a read-only memory, a disk, or an optical disk. All or part of the steps in the above embodiments can also be implemented using one or more integrated circuits. Accordingly, each module in the above embodiments can be implemented in hardware, for example, by implementing its corresponding function through an integrated circuit, or it can be implemented in the form of a software functional module, for example, by a processor executing a program / instruction stored in memory to implement its corresponding function. Embodiments of the present invention are not limited to any particular combination of hardware and software.

[0139] Compared with the prior art, the advantages of the present invention are as follows:

[0140] 1. Compared with traditional tag noise robust systems, it is not limited to a certain noise assumption, but distinguishes and processes noise by pre-identifying the noise type, supports open set noise filtering and closed set noise correction, and reduces the false filtering of closed set noise and the false correction of open set noise while ensuring accuracy.

[0141] 2. This invention addresses the heterogeneity of noise among different clients by calculating corresponding interval thresholds based on different Gaussian mixture distributions of each client, and adaptively guides noise identification for each client.

[0142] 3. This invention utilizes the inherent client (local)-server (global) collaboration characteristic of federated learning to estimate the interval threshold of noise, thereby freeing noise identification from dependence on additional reference distributions, enhancing the feasibility of applying the model to real-world scenarios, and effectively improving the accuracy of federated learning models.

[0143] 4. This invention requires neither the client to send raw data nor the relevant parameters for each sample (due to data privacy concerns), but instead uses an exchange of Gaussian mixture distributions.

[0144] The above embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions also fall within the scope of the present invention, and the patent protection scope of the present invention should be defined by the claims.

Claims

1. A federated learning method robust to mixed noise, characterized in that, include: Step 1: Build a local metric model on the client side, and train the local metric model for a specified number of rounds using local training data to complete the training of the local metric model. Model parameters These are the local metric model parameters sent to the server; these model parameters... The first sample size of the local training data is sent to the server; the local metric model. Using the subjective logistic function as the loss function, the loss function for a given local training data i is: in, Dirichlet distribution The coefficients in , For the activation function ReLU with respect to the feature data of the local metric model in i The output above indicates support for the first... Evidence of class labels , represents the potential of the Dirichlet distribution, n represents the client number, and the subscript b represents the metric model; Step 2, on the server side, use all the model parameters And obtain the global metric model parameters using the first sample size; Step 3: On the client side, using the global metric model parameters and the local training data, obtain the subjective logistic loss value for each local training data, thereby fitting the local Gaussian mixture distribution of all the local training data. Step 4: On the server side, generate local Gaussian mixture discrete distributions from all the local Gaussian mixture distributions, and aggregate them to obtain a global Gaussian mixture distribution. Obtain the KS distance between the local Gaussian mixture discrete distribution and the global Gaussian mixture distribution, and obtain the distinction threshold based on the KS distance. Step 5: The client selects a training dataset from the local training data based on the discrimination threshold. Step 6: Train the local classification model using the training dataset; correct the labels of the local training data using the local classification model; and send the local classification model parameters and the second sample size of the corrected local training data to the server. Step 7: Obtain the global classification model parameters on the server using all the local classification model parameters and the second sample size; Step 8: On the client side, update the local classification model to the global classification model using the global classification model parameters; Repeat steps 6 through 8 until the global classification model of the client converges. Use the global classification model at this point as the final classification model for the client.

2. The federated learning method as described in claim 1, characterized in that, In step 2, the global metric model parameters are obtained using the following formula. : in, This indicates the amount of local training data. This represents the sum of the first sample sizes for all clients. Indicates the client sequence number. Indicates the number of clients.

3. The federated learning method as described in claim 1, characterized in that, In step 4, the distinction threshold includes a first distinction threshold λ. I Second distinction threshold λ II , λ I Let λ be the inflection point value of the KS distance, used to divide the training dataset into a clean subset of data and an open-set noisy subset of data. II The minimum KS distance is used to divide the training dataset into open-set noisy data subsets and closed-set noisy data subsets. in, For discrete global Gaussian mixture distribution The cumulative distribution function, For discrete local Gaussian mixture distribution The conditional probability distribution, .

4. A federated learning system robust to mixed noise, characterized in that, include: The model building module, located on the client side, is used to build a local metric model. This local metric model is trained for a specified number of epochs using local training data to complete the training. Model parameters These are the local metric model parameters sent to the server; these model parameters... The first sample size of the local training data is sent to the server; the local metric model. Using the subjective logistic function as the loss function, the loss function for a given local training data i is: in, Dirichlet distribution The coefficients in , For the activation function ReLU with respect to the feature data of the local metric model in i The output above indicates support for the first... Evidence of class labels , represents the potential of the Dirichlet distribution, n represents the client number, and the subscript b represents the metric model; The metric model aggregation module, set up on the server side, is used to aggregate all parameters of the model. And obtain the global metric model parameters using the first sample size; A local Gaussian mixture distribution generation module is set on the client and is used to obtain the subjective logistic loss value of each local training data using the global metric model parameters and the local training data, thereby fitting the local Gaussian mixture distribution of all the local training data. The threshold calculation module, set on the server, is used to generate local Gaussian mixture discrete distributions from all the local Gaussian mixture distributions, aggregate them to obtain a global Gaussian mixture distribution, obtain the KS distance between the local Gaussian mixture discrete distribution and the global Gaussian mixture distribution, and obtain the distinction threshold based on the KS distance. A noise recognition module, located on the client, is used to select a training dataset from the local training data based on the discrimination threshold. The noise correction module, located on the client, is used to train the local classification model with the training dataset; correct the labels of the local training data with the local classification model; and send the local classification model parameters and the second sample size of the corrected local training data to the server. The classification model aggregation module, set on the server, is used to obtain the global classification model parameters using all local classification model parameters and the second sample size. The model update module, located on the client, is used to update the local classification model to the global classification model using the parameters of the global classification model. The iterative convergence module, set up on the client, is used to repeatedly call the noise correction module, the classification model aggregation module, and the model update module in sequence until the global classification model of the client converges. The global classification model at this point is taken as the final classification model of the client.

5. The federated learning system as described in claim 4, characterized in that, In this metric model aggregation module, the global metric model parameters are obtained using the following formula. : in, This indicates the amount of local training data. This represents the sum of the first sample sizes for all clients. Indicates the client sequence number. Indicates the number of clients.

6. The federated learning system as described in claim 4, characterized in that, The threshold calculation module includes: The first threshold calculation module is used to determine the first distinguishing threshold λ based on the inflection point value of the KS distance. I , λ I This is used to divide the training dataset into a clean subset of data and an open-set noisy subset of data. ; The second threshold calculation module is used to set the minimum value of the KS distance as the second distinguishing threshold λ. II , λ II This is used to divide the training dataset into open-set noisy subsets and closed-set noisy subsets. ; in, For discrete global Gaussian mixture distribution The cumulative distribution function, For discrete local Gaussian mixture distribution The conditional probability distribution, .

7. A computer-readable storage medium storing computer-executable instructions, characterized in that, When the computer-executable instructions are executed, the federated learning method robust to mixed noise as described in any one of claims 1 to 3 is implemented.

8. A data processing apparatus comprising a computer-readable storage medium as claimed in claim 7 and a processor; when the data processing apparatus is used as a client or a server, the processor retrieves and executes computer-executable instructions in the computer-readable storage medium to perform federated learning robust to mixed noise.

Images