Joint learning training method and apparatus

Abstract

The present disclosure relates to the technical field of federated learning, and provides a federated learning training method and device. The method comprises: after a model demander initiates a training task, determining a plurality of participants by a federated learning training center according to the training task; performing federated learning training on a neural network model by using the plurality of participant data of the plurality of participants; obtaining a work log corresponding to each participant in the federated learning training, and obtaining supervision information corresponding to each participant by a security supervision center; and performing auditing processing on the neural network model according to the work log and the supervision information corresponding to each participant, to obtain a trusted federated learning model. By using the above technical means, the problem that the use of participant data in federated learning training cannot be monitored, data security risks exist, and the reliability of federated learning training cannot be guaranteed in the prior art is solved.

Description

Technical Field

[0001] This disclosure relates to the field of collaborative learning technology, and in particular to a collaborative learning training method and apparatus. Background Technology

[0002] In machine learning training, we only have access to the model before and after training. The model during training is like a black box; we cannot understand the model's current state or the training process, such as the parameters of each network layer. Furthermore, because model parameters can be updated using methods like backpropagation, there isn't much research on the model during training. This leads to the inability to monitor training data. If model training centers misuse data from multiple participants, there are data security risks. Additionally, the final model may not meet the requirements of the training task, meaning joint learning training has low reliability.

[0003] In realizing the present invention, the inventors discovered that the related technologies have at least the following technical problems: the use of data by participants in joint learning training cannot be monitored, which poses data security risks and makes it impossible to guarantee the reliability of joint learning training. Summary of the Invention

[0004] In view of this, the present disclosure provides a federated learning training method, apparatus, electronic device, and computer-readable storage medium to solve the problems in the prior art that the use of data by participants in federated learning training cannot be monitored, resulting in data security risks and the inability to guarantee the reliability of federated learning training.

[0005] A first aspect of this disclosure provides a joint learning training method, comprising: after a model requester initiates a training task, determining multiple participants through a joint learning training center based on the training task; jointly training a neural network model using data from the multiple participants; during the joint learning training, obtaining work logs corresponding to each participant and obtaining regulatory information corresponding to each participant through a security monitoring center; and auditing the neural network model based on the work logs and regulatory information corresponding to each participant to obtain a trustworthy joint learning model.

[0006] A second aspect of this disclosure provides a joint learning training apparatus, comprising: a determination module configured to determine multiple participants through a joint learning training center based on the training task initiated by a model requester; a training module configured to perform joint learning training on a neural network model using data from the multiple participants; an acquisition module configured to acquire work logs corresponding to each participant during joint learning training and acquire regulatory information corresponding to each participant through a security monitoring center; and an audit module configured to perform audit processing on the neural network model based on the work logs and regulatory information corresponding to each participant to obtain a trustworthy joint learning model.

[0007] A third aspect of this disclosure provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0008] A fourth aspect of this disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0009] The beneficial effects of this disclosed embodiment compared to the prior art are as follows: After the model requester initiates a training task, multiple participants are identified through a joint learning training center based on the training task; the neural network model is jointly trained using data from multiple participants; during joint learning training, the work logs corresponding to each participant are obtained, and the regulatory information corresponding to each participant is obtained through a security monitoring center; the neural network model is audited based on the work logs and regulatory information corresponding to each participant to obtain a trustworthy joint learning model. By employing the above technical means, the problems of existing technologies, such as the inability to monitor the use of participant data during joint learning training, the existence of data security risks, and the inability to guarantee the reliability of joint learning training, can be solved, thereby avoiding data security risks in joint learning training and improving the reliability of joint learning training. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this disclosure, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of a joint learning architecture according to an embodiment of this disclosure;

[0012] Figure 2 This is a flowchart illustrating a joint learning training method provided in an embodiment of this disclosure;

[0013] Figure 3 This is a schematic diagram of the structure of a collaborative learning training device provided in an embodiment of this disclosure;

[0014] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0015] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, so as to provide a thorough understanding of the embodiments of this disclosure. However, those skilled in the art will understand that this disclosure may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this disclosure with unnecessary detail.

[0016] Federation learning refers to the comprehensive utilization of multiple AI (Artificial Intelligence) technologies, under the premise of ensuring data security and user privacy, to collaboratively explore the value of data and foster new intelligent business forms and models based on joint modeling. Federation learning has at least the following characteristics:

[0017] (1) Participating nodes control their own data in a weakly centralized joint training mode to ensure data privacy and security in the process of co-creating intelligence.

[0018] (2) In different application scenarios, various model aggregation optimization strategies are established by using screening and / or combination of AI algorithms and privacy-preserving computing to obtain high-level and high-quality models.

[0019] (3) Under the premise of ensuring data security and user privacy, based on multiple model aggregation optimization strategies, obtain methods to improve the performance of the federated learning engine. The performance methods can be improved by solving problems such as parallel computing architecture, information interaction under large-scale cross-domain networks, intelligent perception, and anomaly handling mechanisms.

[0020] (4) Obtain the needs of multiple users in various scenarios, determine the true contribution of each joint participant through a mutual trust mechanism, and allocate incentives accordingly.

[0021] Based on the above approach, an AI technology ecosystem based on collaborative learning can be established, fully leveraging the value of industry data and promoting the implementation of scenarios in vertical fields.

[0022] A joint learning training method and apparatus according to embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings.

[0023] Figure 1 This is a schematic diagram of a joint learning architecture according to an embodiment of this disclosure. Figure 1 As shown, the architecture of joint learning may include a server (central node) 101 and participants 102, 103 and 104.

[0024] In the joint learning process, a basic model can be established through server 101, which then sends this model to participants 102, 103, and 104 with whom it has established a communication connection. Alternatively, any participant can establish the basic model and upload it to server 101, which then sends it to other participants with whom it has established a communication connection. Participants 102, 103, and 104 construct models based on the downloaded basic structure and model parameters, train the models using local data, obtain updated model parameters, and encrypt and upload these updated model parameters to server 101. Server 101 aggregates the model parameters sent by participants 102, 103, and 104 to obtain global model parameters, which are then transmitted back to participants 102, 103, and 104. Participants 102, 103, and 104 iterate on their respective models based on the received global model parameters until the models converge, thus achieving model training. During the collaborative learning process, the data uploaded by participants 102, 103, and 104 are model parameters. Local data is not uploaded to server 101, and all participants can share the final model parameters. Therefore, collaborative modeling can be achieved while ensuring data privacy. It should be noted that the number of participants is not limited to the three mentioned above, but can be set as needed. This embodiment of the disclosure does not impose any restrictions on this.

[0025] Figure 2 This is a flowchart illustrating a joint learning training method provided in an embodiment of this disclosure. Figure 2 The joint learning training method can be derived from Figure 1 The server executes this. For example... Figure 2 As shown, the joint learning training method includes:

[0026] S201: After the model requester initiates a training task, multiple participants are identified through the joint learning training center based on the training task.

[0027] S202, using data from multiple stakeholders to jointly train a neural network model;

[0028] S203: In joint learning and training, obtain the work logs of each participant and obtain the supervision information of each participant through the security supervision center;

[0029] S204, audit the neural network model based on the work logs and regulatory information of each participant to obtain a trustworthy joint learning model.

[0030] Existing technologies cannot monitor the use of participant data during joint learning training. This disclosure adds auditing processes to the joint learning training based on the work logs and regulatory information corresponding to each participant, thereby obtaining a trustworthy joint learning model. A trustworthy joint learning model can be understood as one that can be trusted by multiple participants, conforms to the training task, and ensures that the entire joint learning training process complies with regulations, such as not disclosing participant data.

[0031] Auditing the neural network model based on the work logs and regulatory information of each participant can be used as a trusted joint learning model. The neural network model can be any commonly used neural network model, such as Faster-RCNN. Auditing can be understood as processing the neural network model based on the work logs and regulatory information of each participant. This processing can include updating the model parameters or managing interactions between participants and the joint learning training center.

[0032] Joint learning training of a neural network model is performed using data from multiple stakeholders. This includes: for each stakeholder, training a neural network model using their own stakeholder data to obtain their corresponding network model; the stakeholder then uploads the model parameters of their network model to a joint learning training center; the joint learning training center aggregates the model parameters from multiple stakeholders' network models to obtain global parameters; and updates the neural network model's parameters based on these global parameters to obtain the joint learning model. If this joint learning model undergoes auditing, it is considered a trustworthy joint learning model.

[0033] According to the technical solution provided in this disclosure, after the model requester initiates a training task, multiple participants are identified through a joint learning training center based on the training task. The neural network model is then jointly trained using data from these multiple participants. During the joint learning training, work logs for each participant are obtained, and regulatory information for each participant is obtained through a security monitoring center. The neural network model is audited based on the work logs and regulatory information for each participant to obtain a trustworthy joint learning model. By employing the above technical means, the problems of data security risks and the inability to guarantee the reliability of joint learning training in existing technologies can be solved, thereby avoiding data security risks in joint learning training and improving the reliability of joint learning training.

[0034] Each participant's corresponding work log includes: data preprocessing log, local training log, upload log, and download log; the corresponding regulatory information for each participant is the information exchanged between each participant and the joint learning training center, including: information on the participant uploading models to the joint learning training center and information on the participant downloading models from the joint learning training center.

[0035] For a single participant: the data preprocessing log is the packet capture log during the preprocessing of the participant's data. Preprocessing includes: filtering participant data (the participant data may contain data that does not conform to the current training task), etc.; the local training log is the packet capture log during the participant's training process; the upload log is the packet capture log during the process of the participant uploading the model to the joint learning training center; the download log is the packet capture log during the process of the participant downloading the model from the joint learning training center. Uploading the model can be the model parameters of the network model corresponding to the uploader, and downloading the model can be the model parameters of the joint learning model (in a single joint learning training session, there may be multiple training rounds, so there will be multiple uploads and downloads of the model).

[0036] Each participant's work log is the log of packets captured by each participant during the joint learning and training; the corresponding regulatory information for each participant is the information obtained by the security supervision center in supervising each participant during the joint learning and training.

[0037] The data preprocessing log, local training log, upload log, and download log all include: external network communication information, computing resource usage information, model library read / write information, and database read information.

[0038] For a participant: external network communication information refers to the packet capture information during the communication between the participant and the joint learning training center; computing resource usage information refers to the information on the relevant computing hardware usage of the participant during joint learning training; model library read / write information refers to whether the participant reads the model library and writes the model to the model library during joint learning training; database read information refers to whether the participant reads the database during joint learning training.

[0039] For example, if the database read information in the upload log indicates that the participant read the database during the joint learning training, there is a risk of data leakage (the upload log only shows that the participant is uploading the model to the joint learning training center; the participant does not need to upload data to the joint learning training center).

[0040] In step 204, the neural network model is audited based on the work logs and regulatory information corresponding to each participant to obtain a trusted joint learning model. This includes: obtaining the first model parameters from the local training logs of each participant, wherein the work logs of each participant include the local training logs; obtaining the second model parameters from the information on the model uploaded by each participant to the joint learning training center, wherein the regulatory information of each participant includes the information on the model uploaded by the participant to the joint learning training center; performing hash calculations on the first and second model parameters corresponding to each participant to obtain a first hash value and a second hash value; and auditing the neural network model using the security regulatory center based on the first and second hash values ​​corresponding to each participant to obtain a trusted joint learning model.

[0041] The first model parameter for each participant can be the model parameter of the network model trained by each participant; the second model parameter for each participant can be the model parameter uploaded by each participant to the joint learning training center.

[0042] The process involves retrieving the second model parameters from the information uploaded by each participating joint learning training center. This includes obtaining parameter information from the uploaded model information and performing decryption operations such as deserialization on the parameter information to obtain the second model parameters. Serialization is the process of converting a data structure or object into a binary string; serialization can be considered encryption. Deserialization is the process of converting the binary string generated during serialization back into a data structure or object; deserialization can be considered decryption. The model parameters uploaded by each participating joint learning training center have undergone encryption processing such as serialization.

[0043] Hash computation is a one-way cryptographic system, meaning it's an irreversible mapping from plaintext to ciphertext; it only involves encryption and not decryption. Hash computation can prevent data leaks from security monitoring centers.

[0044] Based on the first and second hash values ​​corresponding to each participant, the neural network model is audited by the security monitoring center to obtain a trusted joint learning model. This includes: comparing the first and second hash values ​​corresponding to each participant; when the first and second hash values ​​corresponding to each participant are consistent, the neural network model after joint learning training is determined as a trusted joint learning model; when there is a target participant, a re-upload instruction is sent to the target participant to re-upload the model to the joint learning training center to obtain a trusted joint learning model, wherein the first and second hash values ​​corresponding to the target participant are inconsistent.

[0045] If the first hash value and second hash value are consistent for each participant, it means that the model parameters of the network model trained by each participant are consistent with the model parameters uploaded by each participant to the joint learning training center. Therefore, the neural network model trained through joint learning can be identified as a trustworthy joint learning model. However, if the first hash value and second hash value are inconsistent for some participants (i.e., there are target participants, which can be multiple), it means that the model parameters of the network model trained by some participants are inconsistent with the model parameters uploaded by these participants to the joint learning training center. In this case, the neural network model trained through joint learning cannot be identified as a trustworthy joint learning model. A re-upload instruction should be sent to the target participants to re-upload the model to the joint learning training center.

[0046] In step 204, the neural network model is audited based on the work logs and regulatory information of each participant to obtain a trusted joint learning model. This includes: determining whether each participant has data leakage based on the upload logs of each participant and the information of the model uploaded by each participant to the joint learning training center. The work logs of each participant include upload logs, and the regulatory information of each participant includes the information of the model uploaded by the participant to the joint learning training center. When a participant has data leakage, the security supervision center monitors the interaction between the participant with data leakage and the joint learning training center based on interception rules to obtain a trusted joint learning model. The interception rules include: allowing each participant to interact with the model of the joint learning training center, but disallowing each participant to interact with the data of the joint learning training center.

[0047] Each participant uploads either a model or model parameters to the Joint Learning and Training Center; participants are not allowed to upload participant data. If a participant's upload logs and the information about the model uploaded to the Joint Learning and Training Center indicate that the participant has uploaded participant data, then that participant has committed a data breach. In this case, the interaction between the participant with the data breach and the Joint Learning and Training Center should be monitored; that is, the participant should be allowed to interact with the Joint Learning and Training Center about the model, but not about interacting with the Joint Learning and Training Center about data.

[0048] In step 204, the neural network model is audited based on the work logs and regulatory information corresponding to each participant to obtain a trustworthy joint learning model. This includes: obtaining an alarm rule set, wherein the alarm rule set includes multiple alarm rules, each alarm rule corresponding to a monitoring item in the work logs and / or regulatory information, and the work logs and / or regulatory information include multiple monitoring items; when a target monitoring item in the monitoring log violates the target alarm rule corresponding to the target monitoring item in the alarm rule set, a target alarm is issued; and the joint learning training is processed based on the alarm information corresponding to the target alarm to obtain a trustworthy joint learning model.

[0049] The target monitoring item is a total training epoch of 10. A related monitoring item is the number of interactions between the participant and the training center, which is 15. The target alarm rule is to issue an alarm when the number of interactions exceeds the total training epochs (normally, the number of interactions equals the total training epochs; if the number of interactions exceeds the total training epochs, such as the 11th interaction, the interaction between the participant and the training center is no longer with the model but may be with the participant's data, posing a security risk. In joint learning training, by default, the interaction between the participant and the training center can only be with the model, not with the participant's data). The alarm information corresponding to the target alarm indicates that model training stopped after the 10th epoch, so the neural network model trained after the 11th epoch and beyond is discarded, and the neural network model trained after the 10th epoch is used as the joint learning model.

[0050] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0051] The following are embodiments of the apparatus disclosed herein, which can be used to execute embodiments of the method disclosed herein. For details not disclosed in the apparatus embodiments of this disclosure, please refer to the embodiments of the method disclosed herein.

[0052] Figure 3 This is a schematic diagram of a collaborative learning training device provided in an embodiment of this disclosure. Figure 3 As shown, the joint learning training device includes:

[0053] Module 301 is configured to identify multiple participants based on the training task.

[0054] Training module 302 is configured to jointly train a neural network model using participant data from multiple participants.

[0055] Module 303 is configured to retrieve monitoring logs during joint learning training.

[0056] Processing module 304 is configured to audit the neural network model based on monitoring logs to obtain a trustworthy joint learning model.

[0057] Existing technologies cannot monitor the use of participant data during joint learning training. This disclosure adds auditing processes to the joint learning training based on the work logs and regulatory information corresponding to each participant, thereby obtaining a trustworthy joint learning model. A trustworthy joint learning model can be understood as one that can be trusted by multiple participants, conforms to the training task, and ensures that the entire joint learning training process complies with regulations, such as not disclosing participant data.

[0058] Auditing the neural network model based on the work logs and regulatory information of each participant can be used as a trusted joint learning model. The neural network model can be any commonly used neural network model, such as Faster-RCNN. Auditing can be understood as processing the neural network model based on the work logs and regulatory information of each participant. This processing can include updating the model parameters or managing interactions between participants and the joint learning training center.

[0059] Optionally, the acquisition module 303 is further configured to train a neural network model for each participant using their participant data, ultimately obtaining the network model corresponding to that participant. The participant then uploads the model parameters of the network model to the joint learning training center. The joint learning training center aggregates the model parameters of the network models corresponding to multiple participants to obtain global parameters, and updates the model parameters of the neural network model based on these global parameters to obtain the joint learning model. If this joint learning model undergoes auditing processing, it is considered a trustworthy joint learning model.

[0060] According to the technical solution provided in this disclosure, after the model requester initiates a training task, multiple participants are identified through a joint learning training center based on the training task. The neural network model is then jointly trained using data from these multiple participants. During the joint learning training, work logs for each participant are obtained, and regulatory information for each participant is obtained through a security monitoring center. The neural network model is audited based on the work logs and regulatory information for each participant to obtain a trustworthy joint learning model. By employing the above technical means, the problems of data security risks and the inability to guarantee the reliability of joint learning training in existing technologies can be solved, thereby avoiding data security risks in joint learning training and improving the reliability of joint learning training.

[0061] Each participant's corresponding work log includes: data preprocessing log, local training log, upload log, and download log; the corresponding regulatory information for each participant is the information exchanged between each participant and the joint learning training center, including: information on the participant uploading models to the joint learning training center and information on the participant downloading models from the joint learning training center.

[0062] For a single participant: the data preprocessing log is the packet capture log during the preprocessing of the participant's data. Preprocessing includes: filtering participant data (the participant data may contain data that does not conform to the current training task), etc.; the local training log is the packet capture log during the participant's training process; the upload log is the packet capture log during the process of the participant uploading the model to the joint learning training center; the download log is the packet capture log during the process of the participant downloading the model from the joint learning training center. Uploading the model can be the model parameters of the network model corresponding to the uploader, and downloading the model can be the model parameters of the joint learning model (in a single joint learning training session, there may be multiple training rounds, so there will be multiple uploads and downloads of the model).

[0063] Each participant's work log is the log of packets captured by each participant during the joint learning and training; the corresponding regulatory information for each participant is the information obtained by the security supervision center in supervising each participant during the joint learning and training.

[0064] The data preprocessing log, local training log, upload log, and download log all include: external network communication information, computing resource usage information, model library read / write information, and database read information.

[0065] For a participant: external network communication information refers to the packet capture information during the communication between the participant and the joint learning training center; computing resource usage information refers to the information on the relevant computing hardware usage of the participant during joint learning training; model library read / write information refers to whether the participant reads the model library and writes the model to the model library during joint learning training; database read information refers to whether the participant reads the database during joint learning training.

[0066] For example, if the database read information in the upload log indicates that the participant read the database during the joint learning training, there is a risk of data leakage (the upload log only shows that the participant is uploading the model to the joint learning training center; the participant does not need to upload data to the joint learning training center).

[0067] Optionally, the audit module 304 is further configured to: obtain the first model parameters from the local training logs of each participant, wherein the working logs of each participant include the local training logs; obtain the second model parameters from the information on the model uploaded by each participant to the joint learning training center, wherein the regulatory information of each participant includes the information on the model uploaded by the participant to the joint learning training center; perform hash calculations on the first and second model parameters of each participant to obtain a first hash value and a second hash value; and, based on the first and second hash values ​​of each participant, use the security regulatory center to audit the neural network model to obtain a trustworthy joint learning model.

[0068] The first model parameter for each participant can be the model parameter of the network model trained by each participant; the second model parameter for each participant can be the model parameter uploaded by each participant to the joint learning training center.

[0069] The system retrieves parameter information from the models uploaded by each participating joint learning and training center, and performs decryption operations such as deserialization on these parameters to obtain the second model parameters. Serialization is the process of converting a data structure or object into a binary string; it can be viewed as encryption. Deserialization is the process of converting the binary string generated during serialization back into a data structure or object; it can be viewed as decryption. The model parameters uploaded by each participating joint learning and training center have undergone encryption processing such as serialization.

[0070] Hash computation is a one-way cryptographic system, meaning it's an irreversible mapping from plaintext to ciphertext; it only involves encryption and not decryption. Hash computation can prevent data leaks from security monitoring centers.

[0071] Optionally, the audit module 304 is further configured to compare the first hash value and the second hash value corresponding to each participant: when the first hash value and the second hash value corresponding to each participant are consistent, the neural network model after joint learning training is determined as a trustworthy joint learning model; when there is a target participant, a re-upload instruction is sent to the target participant so that the target participant re-uploads the model to the joint learning training center to obtain a trustworthy joint learning model, wherein the first hash value and the second hash value corresponding to the target participant are inconsistent.

[0072] If the first hash value and second hash value are consistent for each participant, it means that the model parameters of the network model trained by each participant are consistent with the model parameters uploaded by each participant to the joint learning training center. Therefore, the neural network model trained through joint learning can be identified as a trustworthy joint learning model. However, if the first hash value and second hash value are inconsistent for some participants (i.e., there are target participants, which can be multiple), it means that the model parameters of the network model trained by some participants are inconsistent with the model parameters uploaded by these participants to the joint learning training center. In this case, the neural network model trained through joint learning cannot be identified as a trustworthy joint learning model. A re-upload instruction should be sent to the target participants to re-upload the model to the joint learning training center.

[0073] Optionally, the audit module 304 is further configured to determine whether each participant has data leakage based on the upload logs corresponding to each participant and the information of the model uploaded by each participant to the joint learning and training center. The work logs corresponding to each participant include upload logs, and the regulatory information corresponding to each participant includes the information of the model uploaded by that participant to the joint learning and training center. When a participant has data leakage, the security monitoring center monitors the interaction between the participant with data leakage and the joint learning and training center based on interception rules to obtain a trustworthy joint learning model. The interception rules include: allowing each participant to interact with the model of the joint learning and training center, but disallowing each participant to interact with the data of the joint learning and training center.

[0074] Each participant uploads either a model or model parameters to the Joint Learning and Training Center; participants are not allowed to upload participant data. If a participant's upload logs and the information about the model uploaded to the Joint Learning and Training Center indicate that the participant has uploaded participant data, then that participant has committed a data breach. In this case, the interaction between the participant with the data breach and the Joint Learning and Training Center should be monitored; that is, the participant should be allowed to interact with the Joint Learning and Training Center about the model, but not about interacting with the Joint Learning and Training Center about data.

[0075] Optionally, the audit module 304 is further configured to obtain an alarm rule set, wherein the alarm rule set includes multiple alarm rules, each alarm rule corresponding to a monitoring item in the work log and / or regulatory information, and the work log and / or regulatory information includes multiple monitoring items; when a target monitoring item in the monitoring log violates the target alarm rule corresponding to the target monitoring item in the alarm rule set, a target alarm is issued; the joint learning training is processed according to the alarm information corresponding to the target alarm to obtain a trustworthy joint learning model.

[0076] The target monitoring item is a total training epoch of 10. A related monitoring item is the number of interactions between the participant and the training center, which is 15. The target alarm rule is to issue an alarm when the number of interactions exceeds the total training epochs (normally, the number of interactions equals the total training epochs; if the number of interactions exceeds the total training epochs, such as the 11th interaction, the interaction between the participant and the training center is no longer with the model but may be with the participant's data, posing a security risk. In joint learning training, by default, the interaction between the participant and the training center can only be with the model, not with the participant's data). The alarm information corresponding to the target alarm indicates that model training stopped after the 10th epoch, so the neural network model trained after the 11th epoch and beyond is discarded, and the neural network model trained after the 10th epoch is used as the joint learning model.

[0077] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this disclosure.

[0078] Figure 4 This is a schematic diagram of the electronic device 4 provided in an embodiment of this disclosure. Figure 4 As shown, the electronic device 4 of this embodiment includes: a processor 401, a memory 402, and a computer program 403 stored in the memory 402 and executable on the processor 401. When the processor 401 executes the computer program 403, it implements the steps in the various method embodiments described above. Alternatively, when the processor 401 executes the computer program 403, it implements the functions of each module / unit in the various device embodiments described above.

[0079] For example, computer program 403 may be divided into one or more modules / units, which are stored in memory 402 and executed by processor 401 to perform the present disclosure. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 403 in electronic device 4.

[0080] Electronic device 4 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 4 may include, but is not limited to, processor 401 and memory 402. Those skilled in the art will understand that... Figure 4 This is merely an example of electronic device 4 and does not constitute a limitation on electronic device 4. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device may also include input / output devices, network access devices, buses, etc.

[0081] Processor 401 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0082] The memory 402 can be an internal storage unit of the electronic device 4, such as a hard disk or RAM. The memory 402 can also be an external storage device of the electronic device 4, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory 402 can include both internal and external storage units of the electronic device 4. The memory 402 is used to store computer programs and other programs and data required by the electronic device. The memory 402 can also be used to temporarily store data that has been output or will be output.

[0083] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0084] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0085] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.

[0086] In the embodiments provided in this disclosure, it should be understood that the disclosed devices / electronic devices and methods can be implemented in other ways. For example, the device / electronic device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. Multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0087] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0088] Furthermore, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0089] If an integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. A computer-readable medium may include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in a computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0090] The above embodiments are only used to illustrate the technical solutions of this disclosure, and are not intended to limit it. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be included within the protection scope of this disclosure.

Claims

1. A joint learning training method, characterized in that, include: After the model requester initiates a training task, multiple participants are identified through the joint learning training center based on the training task. The neural network model is jointly trained using data from multiple stakeholders. In the joint learning training, the work logs of each participant are obtained, and the supervision information of each participant is obtained through the security supervision center. The neural network model is audited based on the work logs and regulatory information corresponding to each participant to obtain a trustworthy joint learning model; The work logs corresponding to each participant include: data preprocessing logs, local training logs, upload logs, and download logs; The regulatory information corresponding to each participant is the information that each participant interacts with the joint learning and training center, including: information on the participant uploading models to the joint learning and training center and information on the participant downloading models from the joint learning and training center; The data preprocessing log, local training log, upload log, and download log all include: external network communication information, computing resource usage information, model library read / write information, and database read information; The step of auditing the neural network model based on the work logs and regulatory information corresponding to each participant to obtain a trustworthy joint learning model includes: Obtain the first model parameters from the local training logs of each participant, wherein the working logs of each participant include the local training logs; Obtain the second model parameters from the information of the model uploaded by each participating party to the joint learning and training center, wherein the regulatory information corresponding to each participating party includes the information of the model uploaded by that participating party to the joint learning and training center; Hash calculations are performed on the first model parameters and the second model parameters corresponding to each participant to obtain the first hash value and the second hash value; Based on the first hash value and the second hash value corresponding to each participant, the security supervision center performs audit processing on the neural network model to obtain the trusted joint learning model; The step of auditing the neural network model based on the work logs and regulatory information corresponding to each participant to obtain a trustworthy joint learning model further includes: Based on the upload logs corresponding to each participant and the information of the model uploaded by each participant to the joint learning and training center, it is determined whether each participant has data leakage. The work logs corresponding to each participant include the upload logs, and the supervision information corresponding to each participant includes the information of the model uploaded by the participant to the joint learning and training center. When a participant is involved in a data breach, the security monitoring center monitors the interaction between the participant and the joint learning training center based on the interception rules to obtain the trusted joint learning model. The interception rules include: allowing each participant to interact with the model of the joint learning training center, and disallowing each participant to interact with the data of the joint learning training center.

2. The method according to claim 1, characterized in that, The process of auditing the neural network model based on the first and second hash values ​​corresponding to each participant, using the security monitoring center, to obtain a trusted joint learning model includes: Compare the first hash value and the second hash value for each participant: When the first hash value and the second hash value corresponding to each participant are consistent, the neural network model after the joint learning training is determined as the trusted joint learning model. When a target participant exists, a re-upload instruction is sent to the target participant so that the target participant re-uploads the model to the joint learning training center to obtain the trusted joint learning model, wherein the first hash value and the second hash value corresponding to the target participant are inconsistent.

3. The method according to claim 1, characterized in that, The step of auditing the neural network model based on the work logs and regulatory information corresponding to each participant to obtain a trustworthy joint learning model further includes: Obtain an alarm rule set, wherein the alarm rule set includes multiple alarm rules, each alarm rule corresponds to a monitoring item in the work log and / or the regulatory information, and the work log and / or the regulatory information includes multiple monitoring items; When a target monitoring item in the monitoring log violates the target alarm rule corresponding to the target monitoring item in the alarm rule set, a target alarm is issued. The joint learning training is processed based on the alarm information corresponding to the target alarm to obtain a reliable joint learning model.

4. A collaborative learning training device, characterized in that, include: The determination module is configured to identify multiple participants through the joint learning training center based on the training task initiated by the model requester. The training module is configured to jointly train a neural network model using data from multiple stakeholders. The acquisition module is configured to acquire the work logs of each participant during the joint learning training, and to acquire the regulatory information of each participant through the security supervision center. The audit module is configured to audit the neural network model based on the work logs and regulatory information corresponding to each participant, so as to obtain a trustworthy joint learning model. The work logs corresponding to each participant include: data preprocessing logs, local training logs, upload logs, and download logs; the regulatory information corresponding to each participant is the information exchanged between each participant and the joint learning training center, including: information on the participant uploading models to the joint learning training center and information on the participant downloading models from the joint learning training center; the data preprocessing logs, local training logs, upload logs, and download logs all include: external network communication information, computing resource usage information, model library read / write information, and database read information; The audit module is specifically configured to: obtain the first model parameters from the local training logs of each participant, wherein the working logs of each participant include the local training logs; obtain the second model parameters from the information on the model uploaded by each participant to the joint learning training center, wherein the regulatory information of each participant includes the information on the model uploaded by the participant to the joint learning training center; perform hash calculations on the first and second model parameters corresponding to each participant to obtain a first hash value and a second hash value; and, based on the first and second hash values ​​corresponding to each participant, use the security regulatory center to audit the neural network model to obtain the trusted joint learning model. The audit module is further configured to: determine whether each participant has data leakage based on the upload logs corresponding to each participant and the information of the model uploaded by each participant to the joint learning and training center; wherein the work logs corresponding to each participant include the upload logs, and the regulatory information corresponding to each participant includes the information of the model uploaded by the participant to the joint learning and training center; when a participant has data leakage, the security monitoring center monitors the interaction between the participant with data leakage and the joint learning and training center based on the interception rules to obtain the trusted joint learning model; wherein the interception rules include: allowing each participant to interact with the model of the joint learning and training center, and disallowing each participant to interact with the data of the joint learning and training center.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 3.

6. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 3.

Images