Secure computing method based on federated learning

Abstract

The present invention discloses a secure computing method based on joint learning. A joint learning model is established, different data of different objects are imported into the joint learning model, and data features are extracted and stored by a feature extraction network. Two independent but interrelated learning features are trained into two similar models through an encryption algorithm, and one model is used to learn another model to achieve value sharing. In addition, the secure computing method for joint learning of the present invention is based on an encrypted state. When two objects refer to each other's model for learning, they are all privacy operations, which not only protects the data information of both parties, but also provides a confidentiality channel for the joint learning of data of both parties. It has high security value and is worthy of large-scale promotion and application.

Description

Technical Field

[0001] The present invention relates to the technical field of joint learning, and in particular to a secure computing method based on joint learning. Background Art

[0002] A typical example of federated learning is: two things happen repeatedly very close in time, and eventually a connection is gradually formed in the brain. For example, classic conditioning and operant conditioning belong to this type of learning. Existing applications of federated learning refer to the use of a series of algorithms to solve machine learning problems. Their important details are different, but the basic idea is the same: although the server cannot collect data, it can collect model parameters, the server coordinates edge devices to participate in training, and each edge device has training data. Each edge device uses its own data to train a local model, and uploads its own parameters to the server with or without encryption. The server averages or weighted averages the collected parameters and broadcasts them to each edge device. This method is still based on the learning process in the machine state and cannot fully demonstrate the necessity of federated learning.

[0003] Through the search of existing technologies, no benefits brought by the practical application of federated learning were found. Although simple federated learning is easy for the public to understand, deep federated learning often has irregular learning consequences such as data leakage and data theft due to the non-standardization of technology. Summary of the invention

[0004] The purpose of the present invention is to provide a secure computing method based on joint learning. By establishing a joint learning model for data feature extraction, a training platform is also provided for joint learning. The data information is obtained based on secure encryption calculations, has high security, and improves the efficiency of value sharing, so as to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The secure computing method based on joint learning includes the following steps:

[0007] S1: Establish a joint learning model: It consists of an encoder, a communication unit, an integration unit and a classifier.

[0008] S2: receiving, storing, and classifying corresponding data into a database through an encoder, a communication unit, and a classifier, and an integration unit integrates the real-time received data into the database according to data features and marks them as first feature data, second feature data, ... Nth feature data;

[0009] S3: According to the learning task, the input task information is preprocessed, and candidate values ​​are identified according to the regular expression related to the task value. The value is used as the trigger word, and the word examples within a certain window size before and after it are spliced ​​as the input of the joint learning model; the bottom layer of the joint learning model is connected to the word embedding layer, and the position information of the value trigger word is added to the output of the word embedding layer to obtain the final distributed representation for each word example, and then the distributed representation is shared and connected to the subsequent feature extraction network;

[0010] S4: The feature extraction network obtains the semantic type of the numerical information unit through the classifier in the joint learning model, accesses the sequence annotation network to obtain several semantic roles related to the numerical trigger word, merges the output of the specific task, and obtains the extraction result of the numerical information unit. For multiple numerical information units extracted from the input text, the encryption algorithm is used to statistically judge the composite relationship between the numerical information units for training integration;

[0011] S5: Training process: Use deep learning networks to integrate a large amount of crawled encryption-related source code data into the encryption-independent source code crawled numerically to build a basic source code library; then cross-compile and pre-process the basic source code library to obtain a binary code library; extract the statistical features and structural features of the binary encryption algorithm respectively and integrate them to generate an encryption algorithm graph, embed the encryption algorithm graph into the neural network to convert it into a vector form, and judge whether the code is similar by comparing the distance between the vectors. After training, a model capable of judging whether the embedded vectors of the encryption algorithm graph are similar is obtained;

[0012] S6: Integrate the data of similar models to obtain the best learning effect, independently encrypt and store the dissimilar model data as feature data, and update it to the feature database of the joint learning model.

[0013] Furthermore, the encoder in S2 is used to encode the input joint data information into a vector, and then input the encoded vector into the communication unit, and the communication unit determines the interaction vector of the joint data through the feature extraction network; based on the integration unit, the encoding vector of the first weight related to the learning data in the joint data and the interaction vector of the second weight are fused to generate a fusion vector, wherein the second weight is greater than the first weight; the fusion vector is input into the classifier to determine the matching degree related to the learning data.

[0014] Furthermore, S2 also includes the updating of feature data, which receives analog quantities, conditions and / or metadata from the learning task terminal, wherein the learning task terminal is electrically coupled to the communication network, and the analog quantities represent the inferred conclusions calculated based on the final learning task; based on the conditions and / or metadata sent by the final learning task, the suggested updates to the joint learning model from each learning task terminal are allowed or rejected, and when the suggested updates are allowed, the received analog quantities are used to update the joint learning model related to the joint learning data, and the feature data is updated at the same time.

[0015] Furthermore, the first feature data, the second feature data, ... the Nth feature data in S2 are stored independently and are not related to each other.

[0016] Furthermore, the construction steps of the feature extraction network in S3 are as follows:

[0017] S301: Designing a neural network model: Based on the computing and storage resources of the cloud server and the IoT terminal, designing a BranchyNet neural network model including a Trunk part of the cloud and IoT device shared structure and its weights, a Branch part of the IoT device, and a Branch part of the cloud;

[0018] S302: Offline training in the cloud: Use a public dataset to train the BranchyNet model in the cloud to obtain the initial weights of the model;

[0019] S303: Deploy the neural network model: deploy the Trunk (switch) part and the Branch (branch) part to each IoT device and cloud server respectively;

[0020] S304: IoT devices and the cloud collaborate in online joint training: Each IoT device uses the private data continuously generated locally for joint training, continuously optimizes and updates the weights of the Trunk part and the Branch part, and obtains a personalized feature extraction network that can adapt to the usage environment or usage preferences of different IoT devices.

[0021] Furthermore, the semantic types in S4 include: naming, reference, counting, ordinal, measurement, and ratio, and the related semantic roles include: value, modifier, value type, unit, attribute, object, and whole.

[0022] Furthermore, S5 also includes the detection process of the encryption algorithm: using the deep learning network to pre-generate a standard encryption algorithm library, selecting a copy of each encryption algorithm that implements the specification and has determined the specific type, and generating an encryption algorithm graph; then, generating an encryption algorithm graph for the unknown type of encryption algorithm to be detected; finally, embedding the encryption algorithm graph generated by the encryption algorithm to be detected into the trained model, and comparing its vector distance with the embedding of the standard encryption algorithm library in turn; the type of the standard algorithm with the shortest vector distance to the algorithm to be detected is the type of the encryption algorithm to be detected.

[0023] Furthermore, the established joint learning model can also be used for enhanced training of data sets. For each piece of data involved in the joint training, there is a certain number of samples. For each of them, joint training is equivalent to enhancing the data set, that is, horizontal joint training; the aggregation of data features is vertical joint training.

[0024] Furthermore, the established joint learning model can also be used for data labeling. Among the participating data items, one has features and the other has labeled data. The two can jointly learn the connection between each other without leaking privacy, thereby achieving the labeling effect.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. The secure computing method based on joint learning provided by the present invention establishes a joint learning model, imports different data of different objects into the joint learning model, and extracts and stores data features by a feature extraction network. Two independent but interrelated learning features are used to train two similar models through an encryption algorithm, and one model is used to learn another model to achieve value sharing.

[0027] 2. The secure computing method based on joint learning provided by the present invention is carried out in an encrypted state. When two objects refer to each other's model for learning, they are all privacy operations. This not only protects the data information of both parties, but also provides a confidential channel for the joint learning of data of both parties. It has high security value and is worthy of large-scale promotion and application.

[0028] 3. The secure computing method based on joint learning provided by the present invention can combine the data samples of each party in the joint learning through the method of the present invention, enrich the statistical accuracy of the joint model, improve the versatility of the model while not leaking each other's sample information.

[0029] 4. The secure computing method based on joint learning provided by the present invention has feature data and the other party has labeled data in each party participating in the joint learning. Through joint computing, both parties can learn the connection between the two without disclosing data privacy to each other, thereby achieving the labeling effect.

[0030] 5. The secure computing method based on joint learning provided by the present invention, in which the multiple parties participating in the joint learning each contain unique partial features, can be used to train a joint model containing multiple features that cannot be completed by a single participant.

[0031] 6. The secure computing method based on joint learning provided by the present invention is aimed at a situation where one party's model has very high value and is unwilling to share, while the other party has private data but is also unwilling to share the data; the present invention can use one party's model to consume the other party's data, but the two do not share with each other, thereby achieving the ultimate goal of jointly generating value but protecting privacy. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a flow chart of the method of the present invention;

[0033] Figure 2 Constructing a graph for the feature extraction network of the present invention;

[0034] Figure 3 is a method flow chart of Embodiment 1 of the present invention;

[0035] Figure 4 This is a method flow chart of Embodiment 2 of the present invention;

[0036] Figure 5 This is a method flow chart of Embodiment 3 of the present invention;

[0037] Figure 6 This is a method flow chart of embodiment 4 of the present invention. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] See also Figure 1-2 In an embodiment of the present invention, there is provided: a secure computing method based on joint learning, comprising the following steps:

[0040] Step 1: Establish a joint learning model: It consists of an encoder, a communication unit, an integration unit and a classifier.

[0041] Step 2: receiving, storing, and classifying the corresponding data into a database through the encoder, the communication unit, and the classifier, and the integration unit integrates the real-time received data into the database according to the data features and marks them as first feature data, second feature data, ... Nth feature data;

[0042] Step 3: According to the learning task, the input task information is preprocessed, and candidate values ​​are identified according to the regular expression related to the task value. The value is used as the trigger word, and the word examples within a certain window size before and after it are spliced ​​as the input of the joint learning model; the bottom layer of the joint learning model is connected to the word embedding layer, and the position information of the value trigger word is added to the output of the word embedding layer to obtain the final distributed representation for each word example, and then the distributed representation is shared and connected to the subsequent feature extraction network;

[0043] Step 4: The feature extraction network obtains the semantic type of the numerical information unit through the classifier in the joint learning model, accesses the sequence annotation network to obtain several semantic roles related to the numerical trigger word, merges the output of the specific task, and obtains the extraction result of the numerical information unit. For multiple numerical information units extracted from the input text, the encryption algorithm is used to statistically judge the composite relationship between the numerical information units for training integration;

[0044] Step 5: Training process: Use deep learning networks to integrate a large amount of encryption-related source code data into the encryption-independent source code crawled numerically to build a basic source code library; then cross-compile and pre-process the basic source code library to obtain a binary code library; extract the statistical features and structural features of the binary encryption algorithm respectively, integrate them to generate an encryption algorithm graph, embed the encryption algorithm graph into the neural network to convert it into a vector form, and judge whether the code is similar by comparing the distance between the vectors. After training, a model capable of judging whether the embedded vectors of the encryption algorithm graph are similar is obtained;

[0045] Step 6: Integrate the data of similar models to obtain the best learning effect, independently encrypt and store the dissimilar model data as feature data, and update it to the feature database of the joint learning model.

[0046] In the above embodiment, in the second step, the encoder is used to encode the input joint data information into a vector, and then input the encoded vector into the communication unit, and the communication unit determines the interaction vector of the joint data through the feature extraction network; based on the integration unit, the encoding vector of the first weight related to the learning data in the joint data and the interaction vector of the second weight are fused to generate a fusion vector, wherein the second weight is greater than the first weight; the fusion vector is input into the classifier to determine the matching degree related to the learning data.

[0047] In the above embodiment, the second step also includes updating the feature data, which receives analog quantities, conditions and / or metadata from the learning task terminal, wherein the learning task terminal is electrically coupled to the communication network, and the analog quantities represent inference conclusions calculated based on the final learning task; based on the conditions and / or metadata sent by the final learning task, the suggested updates to the joint learning model from each learning task terminal are allowed or rejected; when the suggested updates are allowed, the received analog quantities are used to update the joint learning model related to the joint learning data, and the feature data is updated at the same time; wherein the first feature data, the second feature data, ... the Nth feature data are stored independently and are not associated with each other.

[0048] In the above embodiment, the steps of constructing the feature extraction network in the third step are as follows:

[0049] S301: Designing a neural network model: Based on the computing and storage resources of the cloud server and the IoT terminal, designing a BranchyNet neural network model including a Trunk part of the cloud and IoT device shared structure and its weights, a Branch part of the IoT device, and a Branch part of the cloud;

[0050] S302: Offline training in the cloud: Use a public dataset to train the BranchyNet model in the cloud to obtain the initial weights of the model;

[0051] S303: Deploy the neural network model: deploy the Trunk (switch) part and the Branch (branch) part to each IoT device and cloud server respectively;

[0052] S304: IoT devices and the cloud collaborate in online joint training: Each IoT device uses the private data continuously generated locally for joint training, continuously optimizes and updates the weights of the Trunk part and the Branch part, and obtains a personalized feature extraction network that can adapt to the usage environment or usage preferences of different IoT devices.

[0053] The construction and application of the above feature extraction network can meet the needs of a variety of different application scenarios.

[0054] In the above embodiment, the semantic types in the fourth step include: naming, reference, counting, ordinal, measurement and ratio, and the related semantic roles include: value, modification, value type, unit, attribute, object and whole, which are used for feature extraction of the feature extraction network.

[0055] In the above embodiment, the fifth step also includes a detection process for the encryption algorithm: a standard encryption algorithm library is pre-generated using a deep learning network, and a copy of each encryption algorithm that implements the specification and has been determined to be of a specific type is selected, and an encryption algorithm graph is generated; then, an encryption algorithm graph is also generated for the unknown type of encryption algorithm to be detected; finally, the encryption algorithm graph generated by the encryption algorithm to be detected is embedded in the trained model, and its vector distance is compared with the embedding of the standard encryption algorithm library in turn; the type of the standard algorithm with the shortest vector distance to the algorithm to be detected is the type of the encryption algorithm to be detected, and the training process is further verified by the detection process to improve the accuracy and security of the encryption algorithm.

[0056] In order to further better explain the present invention, the following specific embodiments are also provided:

[0057] Embodiment 1:

[0058] See also Figure 3 , the secure computing method based on joint learning of the present invention is applied to data set enhancement training in joint learning, which includes the following steps:

[0059] Step 1: Establish a joint learning model: It consists of an encoder, a communication unit, an integration unit and a classifier.

[0060] Step 2: For each object participating in the joint, a certain number of samples are obtained. The samples are received, stored, and classified into a database through an encoder, a communication unit, and a classifier. The integration unit integrates the real-time received data into the database according to the data features and marks them as the first feature data, the second feature data, ... the Nth feature data;

[0061] Step 3: According to the joint training task, the input task information is preprocessed, and candidate values ​​are identified according to the regular expression related to the task value. The value is used as the trigger word, and the word examples within a certain window size before and after it are spliced ​​as the input of the joint learning model; the bottom layer of the joint learning model is connected to the word embedding layer, and the position information of the value trigger word is added to the output of the word embedding layer to obtain the final distributed representation for each word example, and then the distributed representation is shared and connected to the subsequent feature extraction network;

[0062] Step 4: The feature extraction network obtains the semantic type of the numerical information unit through the classifier in the joint learning model, accesses the sequence annotation network to obtain several semantic roles related to the numerical trigger word, merges the output of the specific task, and obtains the extraction result of the numerical information unit. For multiple numerical information units extracted from the input text, the encryption algorithm is used to statistically judge the composite relationship between the numerical information units for training integration;

[0063] Step 5: Training process: Use deep learning network to integrate a large amount of encryption-related source code data crawled into encryption-independent source code crawled numerically to build a basic source code library; then cross-compile and pre-process the basic source code library to obtain a binary code library; extract statistical features and structural features of the binary encryption algorithm respectively, integrate to generate an encryption algorithm graph, embed the encryption algorithm graph into the neural network to become a vector form, and judge whether the code is similar by comparing the distance between the vectors. After training, a model capable of judging whether the embedded vectors of the encryption algorithm graph are similar is obtained; since each object participating in the joint participation in this embodiment has an independent number of samples, the trained model is dissimilar;

[0064] Step 6: Independently encrypt and store dissimilar model data in the joint learning data as feature data, and update it to the feature database of the joint learning model, which is equivalent to horizontal federation, that is, enhancing the training of the data set.

[0065] Embodiment 2:

[0066] See also Figure 4 , the secure computing method based on joint learning of the present invention is applied to feature aggregation of data sets in joint learning, which includes the following steps:

[0067] Step 1: Establish a joint learning model: It consists of an encoder, a communication unit, an integration unit and a classifier.

[0068] Step 2: For each object participating in the joint, a certain number of samples are obtained. The samples are received, stored, and classified into a database through an encoder, a communication unit, and a classifier. The integration unit integrates the real-time received data into the database according to the data features and marks them as the first feature data, the second feature data, ... the Nth feature data;

[0069] Step 3: According to the joint training task, the input task information is preprocessed, and candidate values ​​are identified according to the regular expression related to the task value. The value is used as the trigger word, and the word examples within a certain window size before and after it are spliced ​​as the input of the joint learning model; the bottom layer of the joint learning model is connected to the word embedding layer, and the position information of the value trigger word is added to the output of the word embedding layer to obtain the final distributed representation for each word example, and then the distributed representation is shared and connected to the subsequent feature extraction network;

[0070] Step 4: The feature extraction network obtains the semantic type of the numerical information unit through the classifier in the joint learning model, accesses the sequence annotation network to obtain several semantic roles related to the numerical trigger word, merges the output of the specific task, and obtains the extraction result of the numerical information unit. For multiple numerical information units extracted from the input text, the encryption algorithm is used to statistically judge the composite relationship between the numerical information units for training integration;

[0071] Step 5: Training process: Use deep learning network to integrate a large amount of encryption-related source code data crawled into the encryption-independent source code crawled numerically to build a basic source code library; then cross-compile and pre-process the basic source code library to obtain a binary code library; extract the statistical features and structural features of the binary encryption algorithm respectively to integrate and generate an encryption algorithm graph, embed the encryption algorithm graph into the neural network into a vector form, and judge whether the code is similar by comparing the distance between the vectors. After training, a model capable of judging whether the embedded vectors of the encryption algorithm graph are similar is obtained; since each object participating in the joint participation in this embodiment has an independent number of samples, the trained models are not similar, but the corresponding feature data can be extracted;

[0072] Step 6: The feature data in the joint learning data is encrypted and stored independently, and updated to the feature database of the joint learning model, which is equivalent to vertical federation, that is, the feature data set of the joint learning model is enhanced.

[0073] Embodiment three:

[0074] See also Figure 5 , the secure computing method based on joint learning of the present invention is applied to data labeling in joint learning, which includes the following steps:

[0075] Step 1: Establish a joint learning model: It consists of an encoder, a communication unit, an integration unit and a classifier.

[0076] Step 2: Each object participating in the joint project has a certain number of samples, some of which are feature data samples and some of which are labeled data samples. The samples are received, stored, and classified into a database through an encoder, a communication unit, and a classifier. The integration unit integrates the real-time received data in the database according to the data features and marks them as the first feature data, the second feature data, ... the Nth feature data;

[0077] Step 3: According to the joint training task, the input task information is preprocessed, and candidate values ​​are identified according to the regular expression related to the task value. The value is used as the trigger word, and the word examples within a certain window size before and after it are spliced ​​as the input of the joint learning model; the bottom layer of the joint learning model is connected to the word embedding layer, and the position information of the value trigger word is added to the output of the word embedding layer to obtain the final distributed representation for each word example, and then the distributed representation is shared and connected to the subsequent feature extraction network;

[0078] Step 4: The feature extraction network obtains the semantic type of the numerical information unit through the classifier in the joint learning model, accesses the sequence annotation network to obtain several semantic roles related to the numerical trigger word, merges the output of the specific task, and obtains the extraction result of the numerical information unit. For multiple numerical information units extracted from the input text, the encryption algorithm is used to statistically judge the composite relationship between the numerical information units for training integration;

[0079] Step 5: Training process: Use a deep learning network to integrate a large amount of crawled encryption-related source code data into the encryption-independent source code that is numerically crawled to build a basic source code library; then cross-compile and preprocess the basic source code library to obtain a binary code library; extract the statistical features and structural features of the binary encryption algorithm respectively, integrate them to generate an encryption algorithm graph, embed the encryption algorithm graph into the neural network into a vector form, and judge whether the codes are similar by comparing the distance between the vectors. After training, a model capable of judging whether the embedded vectors of the encryption algorithm graph are similar is obtained; since some of the objects jointly participating in this embodiment have feature data samples and some have labeled data samples, due to the encryption processing of the joint learning model, the two can learn the connection between them through the joint learning model without leaking privacy, thereby achieving the labeling effect.

[0080] Specific application example: For example, in a medical scenario, one party has the patient's characteristics, and the other party has the patient's physical examination results. Through the secure computing method based on joint learning in this embodiment, data labeling can be jointly performed.

[0081] Embodiment 4:

[0082] See also Figure 6 , the secure computing method based on joint learning of the present invention is applied to the study of model and data independence in joint learning, which includes the following steps:

[0083] Step 1: Establish a joint learning model: It consists of an encoder, a communication unit, an integration unit and a classifier.

[0084] Step 2: For each object participating in the joint, some have models and some have data, which is equivalent to theoretical data and practical data, and is held between the two objects. The data samples are received, stored, and classified into the database through the encoder, communication unit and classifier. The integration unit integrates the real-time received data in the database according to the data features and marks them as the first feature data, the second feature data, ... the Nth feature data;

[0085] Step 3: According to the joint training task, the input task information is preprocessed, and candidate values ​​are identified according to the regular expression related to the task value. The value is used as the trigger word, and the word examples within a certain window size before and after it are spliced ​​as the input of the joint learning model; the bottom layer of the joint learning model is connected to the word embedding layer, and the position information of the value trigger word is added to the output of the word embedding layer to obtain the final distributed representation for each word example, and then the distributed representation is shared and connected to the subsequent feature extraction network;

[0086] Step 4: The feature extraction network obtains the semantic type of the numerical information unit through the classifier in the joint learning model, accesses the sequence annotation network to obtain several semantic roles related to the numerical trigger word, merges the output of the specific task, and obtains the extraction result of the numerical information unit. For multiple numerical information units extracted from the input text, the encryption algorithm is used to statistically judge the composite relationship between the numerical information units for training integration;

[0087] Step 5: Training process: Use deep learning network to integrate a large amount of encryption-related source code data crawled into encryption-independent source code crawled numerically to build a basic source code library; then cross-compile and preprocess the basic source code library to obtain a binary code library; extract the statistical features and structural features of the binary encryption algorithm respectively, integrate them to generate an encryption algorithm graph, embed the encryption algorithm graph into the neural network to become a vector form, and judge whether the code is similar by comparing the distance between the vectors. After training, a model capable of judging whether the embedded vectors of the encryption algorithm graph are similar is obtained; since the two objects jointly participating in this embodiment have the same learning object data, but the data is divided into model data and theoretical data, the two are independent of each other, and can be combined in a general database to train a similar model framework;

[0088] Step 6: In a similar model framework, use one party’s model to consume the other party’s data, achieve value sharing while protecting the privacy of the data model, avoid learning alone, and thus improve the efficiency of joint learning.

[0089] To sum up: the secure computing method based on joint learning provided by the present invention establishes a joint learning model, imports different data of different objects into the joint learning model, and uses a feature extraction network to extract and store data features. The learning features of two independent and different objects but interrelated are trained into two similar models through an encryption algorithm, and one model is used to learn another model to achieve value sharing. In addition, the secure computing method based on joint learning of the present invention is based on an encrypted state. When two objects refer to each other's model for learning, they are all privacy operations. This not only protects the data information of both parties, but also provides a confidentiality channel for the joint learning of data of both parties. It has high security value and is worthy of large-scale promotion and application.

[0090] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A secure computing method based on joint learning, characterized in that: The following steps are involved: S1: Establish a joint learning model: It consists of an encoder, a communication unit, an integration unit, and a classifier; S2: receiving, storing, and classifying corresponding data into a database through an encoder, a communication unit, and a classifier, and an integration unit integrates the real-time received data into the database according to data features and marks them as first feature data, second feature data, ... Nth feature data; S3: According to the learning task, the input task information is preprocessed, and candidate values ​​are identified according to the regular expression related to the task value. The value is used as the trigger word, and the word examples within a certain window size before and after it are spliced ​​as the input of the joint learning model; the bottom layer of the joint learning model is connected to the word embedding layer, and the position information of the value trigger word is added to the output of the word embedding layer to obtain the final distributed representation for each word example, and then the distributed representation is shared and connected to the subsequent feature extraction network; S4: The feature extraction network obtains the semantic type of the numerical information unit through the classifier in the joint learning model, accesses the sequence annotation network to obtain several semantic roles related to the numerical trigger word, merges the output of the specific task, and obtains the extraction result of the numerical information unit. For multiple numerical information units extracted from the input text, the encryption algorithm is used to statistically judge the composite relationship between the numerical information units for training integration; S5: Training process: Use deep learning networks to integrate a large amount of crawled encryption-related source code data into the encryption-independent source code crawled numerically to build a basic source code library; Then, the basic source code library is cross-compiled and preprocessed to obtain a binary code library. The statistical features and structural features of the binary encryption algorithm are extracted and integrated to generate an encryption algorithm graph. The encryption algorithm graph is embedded into the neural network and converted into a vector form. By comparing the distance between the vectors, it is possible to determine whether the codes are similar. After training, a model capable of determining whether the embedded vectors of the encryption algorithm graph are similar is obtained. S6: Integrate the data of similar models to obtain the best learning effect, independently encrypt and store the dissimilar model data as feature data, and update it to the feature database of the joint learning model.

2. The secure computing method based on joint learning according to claim 1, characterized in that: The encoder in S2 is used to encode the input joint data information into a vector, and then input the encoded vector into the communication unit, which determines the interaction vector of the joint data through a feature extraction network; based on the integration unit, the encoding vector of the first weight related to the learning data in the joint data and the interaction vector of the second weight are fused to generate a fusion vector, wherein the second weight is greater than the first weight; the fusion vector is input into the classifier to determine the matching degree related to the learning data.

3. The secure computing method based on joint learning as claimed in claim 2, characterized in that: S2 also includes the updating of feature data, which receives analog quantities, conditions and / or metadata from the learning task terminal, wherein the learning task terminal is electrically coupled to the communication network, and the analog quantities represent the inferred conclusions calculated based on the final learning task; based on the conditions and / or metadata sent by the final learning task, the suggested updates to the joint learning model from each learning task terminal are allowed or rejected; if the suggested updates are allowed, the received analog quantities are used to update the joint learning model related to the joint learning data, and the feature data is updated at the same time.

4. The secure computing method based on joint learning as claimed in claim 3, characterized in that: The first feature data, the second feature data, ... the Nth feature data in S2 are stored independently and are not related to each other.

5. The secure computing method based on joint learning according to claim 1, characterized in that: The steps for constructing the feature extraction network in S3 are as follows: S301: Design a neural network model: Based on the computing and storage resources of the cloud server and IoT terminals, design a BranchyNet neural network model including the Trunk part of the cloud and IoT device shared structure and its weights, the Branch part of the IoT device and the Branch part of the cloud; S302: Offline training in the cloud: Use a public dataset to train the BranchyNet model in the cloud to obtain the initial weights of the model; S303: Deploy the neural network model: deploy the Trunk part and the Branch part to each IoT device and cloud server respectively; S304: IoT devices and the cloud collaborate in online joint training: Each IoT device uses the private data continuously generated locally for joint training, continuously optimizes and updates the weights of the Trunk and Branch parts, and obtains a personalized feature extraction network that can adapt to the usage environment or usage preferences of different IoT devices.

6. The secure computing method based on joint learning according to claim 1, characterized in that: The semantic types in S4 include: naming, reference, counting, ordinal, measurement, and ratio, and the related semantic roles include: value, modifier, value type, unit, attribute, object, and whole.

7. The secure computing method based on joint learning according to claim 1, characterized in that: S5 also includes the encryption algorithm detection process: using the deep learning network to pre-generate a standard encryption algorithm library, selecting a copy of each encryption algorithm that implements the specification and has been determined to be of a specific type, and generating an encryption algorithm graph; then, generating an encryption algorithm graph for the unknown type of encryption algorithm to be detected; finally, embedding the encryption algorithm graph generated by the encryption algorithm to be detected into the trained model, and comparing its vector distance with the embedding of the standard encryption algorithm library in turn; the type of the standard algorithm with the shortest vector distance to the algorithm to be detected is the type of the encryption algorithm to be detected.

8. The secure computing method based on joint learning according to claim 1, characterized in that: The established joint learning model can also be used for data set enhancement training. For each data involved in the joint training, there are a certain number of samples. For each data involved, joint training is equivalent to data set enhancement, that is, horizontal joint training. Its data features are aggregated, that is, vertically combined.

9. The secure computing method based on joint learning according to claim 1, characterized in that: The established joint learning model can also be used for data labeling. Among the participating data items, one has features and the other has labeled data. The two can jointly learn the connection between each other without leaking privacy, thereby achieving the labeling effect.

Images