Federal learning privacy protection method based on proxy data set

Abstract

The invention discloses a federated learning privacy protection method based on a proxy data set, which is characterized in that agreements and multi-device participants reach consensus before federated learning, key elements of each link are clarified, the process is ensured to be standard and consistent, a foundation is laid for privacy protection and efficient training, and original data privacy is protected by constructing the proxy data set and a model. Data characteristics are simulated to ensure effective training, flexibility and adaptability are improved through proxy model optimization of participants, calculation complexity is reduced through logic optimization, training is accelerated, bandwidth consumption is reduced, efficiency is improved, communication security is ensured through establishment of public key infrastructure, data confidentiality, integrity and security audit are realized, and the method is suitable for popularization and application. By formulating an auditing strategy, data compliance and consistency are ensured, third-party auditing verification data are true and effective, federal learning transparency and credibility are enhanced, and safe and efficient data collaboration and learning are promoted.

Description

Technical Field

[0001] The present invention relates to the field of privacy computing technology, and in particular to a method for protecting privacy in federated learning based on a proxy data set. Background Art

[0002] Federated learning is one of the key technologies to achieve secure data circulation. Federated learning is a machine learning method that allows multiple parties to jointly train a shared model while keeping the data decentralized and private. It addresses the need for data security when data cannot be transferred to a central server due to privacy issues, legal restrictions, or large amounts of data. From the perspective of privacy protection, federated learning enables participants to train machine learning models using local data without sharing data with a central server, which helps protect sensitive information such as personal data because the data remains on the data owner's device or server. From the perspective of data ownership and control, through federated learning, data owners retain full ownership and control of their data. They can choose to participate in the collaborative training process while maintaining control over how their data is used and ensuring compliance with privacy regulations. From the perspective of shared security, federated learning allows participants to securely share knowledge while keeping the underlying data private. Participants only exchange model updates or aggregated gradients instead of sharing raw data, which reduces the risk of data leakage or unauthorized access to sensitive information. From the perspective of computational scalability, federated learning utilizes computing power and data resources distributed across multiple devices or servers. This decentralized approach enables efficient training on large-scale data sets without the need for data integration. This reduces network bandwidth requirements and improves scalability. In addition, federated learning also improves model accuracy to some extent because federated learning trains models on diverse and distributed data sets. Federated learning is applied to different industries and fields. For example, in the medical field, federated learning enables multiple healthcare service providers to jointly train machine learning models and use data for collaboration while protecting patient privacy. It can be used for tasks such as disease prediction, medical imaging analysis, drug development, and personalized treatment recommendations. In the field of financial services, federated learning can promote collaborative model training among financial institutions while ensuring data privacy. It can be used for tasks such as fraud detection, credit scoring, risk assessment, and customer behavior analysis without disclosing sensitive financial information.

[0003] However, the trained machine learning model itself may have privacy leakage security issues. Due to the complexity and difficulty of interpreting the machine learning model itself, the machine learning model may remember some additional feature information about the training set during the training process. Usually, this information is sensitive. If leaked, it will inevitably lead to serious privacy security issues. Therefore, it has attracted widespread attention from scholars at home and abroad. For machine learning models, attackers use specific attack strategies to extract hidden privacy information in the model under white-box conditions or black-box conditions of the model prediction API, or infer sensitive information contained in the model training process. Summary of the invention

[0004] In view of this, the present invention proposes a privacy protection method for federated learning based on a proxy data set, which can achieve efficient federated learning training while protecting data privacy.

[0005] The technical solution of the present invention is achieved in this way:

[0006] A method for privacy protection of federated learning based on proxy datasets, specifically including:

[0007] Multiple device participants reach an agreement before participating in training;

[0008] Constructing a proxy data set according to the protocol and training it to form a proxy model;

[0009] Setting the learning tasks and calculation strategies of the agent model;

[0010] Each device participant performs customized optimization of learning tasks and strategies based on the learning tasks and computing strategies of the agent model;

[0011] Optimizing the proxy model using a logic optimization method;

[0012] Establishing a public key infrastructure for identifying participants;

[0013] An audit policy is established, wherein the audit policy is used to audit the data input by each device participant.

[0014] As a further optional solution of the method for protecting privacy of federated learning based on a proxy dataset, the step of constructing a proxy dataset and training a proxy model according to the protocol specifically includes:

[0015] Obtain original data sets from each device participant;

[0016] Determine the sample space of the original data set and extract features covering the sample space;

[0017] Create a virtual data set through a noise generating function;

[0018] Mix the original dataset with the virtual dataset to form a proxy dataset;

[0019] According to the protocol content, the proxy model is trained using the proxy dataset and loss function.

[0020] As a further optional solution of the privacy protection method of federated learning based on proxy datasets, each device participant performs customized optimization of learning tasks and strategies according to the learning tasks and computing strategies of the proxy model, specifically including:

[0021] Obtain learning tasks and strategies customized by each device participant;

[0022] Generate DSL code for each device participant based on the learning tasks and strategies customized by each device participant;

[0023] Call the API interface to execute the customized learning tasks and strategies of each device participant, and optimize the learning tasks and computing strategies of the agent model.

[0024] As a further optional solution of the privacy protection method for federated learning based on proxy datasets, the optimization of the proxy model by means of logic optimization specifically includes:

[0025] Adopt gradient compression technology to optimize the aggregation algorithm of proxy model;

[0026] Distributed computing resources are utilized and parallel computing technology is adopted to accelerate the aggregate update of agent model parameters.

[0027] As a further optional solution of the privacy protection method for federated learning based on proxy datasets, the establishment of a public key infrastructure, which is used to identify the identity of participants, specifically includes:

[0028] Each participant is assigned a unique public and private key pair for encryption and decryption of communication data;

[0029] The identity of the participants is verified through a digital certificate mechanism to ensure the authenticity of the identities of both communicating parties and the integrity of the data.

[0030] As a further optional solution of the privacy protection method for federated learning based on proxy datasets, the establishment of an audit strategy, which is used to audit the data input by each device participant, specifically includes:

[0031] Each participant calculates the committed amount of their data and generates a signature using a private key;

[0032] All parties run secure computations encoding the original learning task and check the consistency of the input data;

[0033] Verify the input data of each participant based on the public key infrastructure (PKI) and detect potential malicious participants.

[0034] A privacy protection system for federated learning based on proxy datasets, comprising:

[0035] The agreement module is used for multiple device participants to reach an agreement before participating in training;

[0036] A proxy data set construction and model training module, used to construct a proxy data set according to the protocol and train to form a proxy model;

[0037] A setting module, used to set the learning tasks and calculation strategies of the agent model;

[0038] The custom optimization module is used by each device participant to perform custom optimization of learning tasks and strategies based on the learning tasks and computing strategies of the agent model;

[0039] A logic optimization module, used for optimizing the proxy model by using a logic optimization method;

[0040] A public key infrastructure module, used for establishing a public key infrastructure, wherein the public key infrastructure is used for identifying the identity of a participant;

[0041] The audit policy module is used to establish an audit policy, and the audit policy is used to audit the data input by each device participant.

[0042] A computing device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above-mentioned method for privacy protection of federated learning based on a proxy data set are implemented.

[0043] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-mentioned method for privacy protection of federated learning based on a proxy data set.

[0044] The beneficial effects of the present invention are as follows: through the protocol, multiple device participants can reach a consensus before participating in the training, clarify key elements such as data preprocessing, model training algorithm, learning tasks, computing strategies, and audit rules in the post-processing stage, which ensures the standardization and consistency of the federated learning process, and lays the foundation for subsequent privacy protection and efficient training. Constructing a proxy data set and training to form a proxy model are key steps to protect the privacy of the original data, so that the original data will not be directly exposed during the training process, thereby effectively preventing the risk of data leakage. At the same time, the proxy model can simulate the characteristics of the original data to a certain extent, ensuring the effectiveness and accuracy of the training. Each device participant performs customized optimization based on the learning tasks and computing strategies of the proxy model, which improves the flexibility and adaptability of federated learning. The proxy model is optimized using a logical optimization method, which can It significantly reduces computational complexity, improves model training speed, realizes efficient federated learning training, reduces network bandwidth consumption, improves data processing efficiency, establishes a public key infrastructure to identify participants, ensures the security of communication data, and establishes a public key infrastructure to assign a unique public and private key pair to each participant, thereby achieving data confidentiality and integrity protection. At the same time, the establishment of a public key infrastructure also supports security audit functions, providing a reliable basis for audits in the post-processing stage, and establishing an audit strategy to audit the data input by each device participant, ensuring data compliance and consistency. Through the audit process, third-party auditors can verify whether the data input by each party complies with the protocol provisions and whether the model update is trained based on real and valid data, which further enhances the transparency and credibility of the federated learning process. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0046] Figure 1 A schematic diagram of the steps of a method for privacy protection of federated learning based on a proxy dataset provided by an embodiment of the present invention;

[0047] Figure 2 A schematic diagram of the composition of a federated learning privacy protection system based on a proxy dataset provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present invention are described clearly and completely below. 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.

[0049] refer to Figure 1 to Figure 2 , a privacy protection method for federated learning based on proxy datasets, specifically including:

[0050] Multiple device participants reach an agreement before participating in training;

[0051] Constructing a proxy data set according to the protocol and training it to form a proxy model;

[0052] Setting the learning tasks and calculation strategies of the agent model;

[0053] Each device participant performs customized optimization of learning tasks and strategies based on the learning tasks and computing strategies of the agent model;

[0054] Optimizing the proxy model using a logic optimization method;

[0055] Establishing a public key infrastructure for identifying participants;

[0056] An audit policy is established, wherein the audit policy is used to audit the data input by each device participant.

[0057] In this embodiment, through the protocol, multiple device participants can reach a consensus before participating in the training, clarify key elements such as data preprocessing, model training algorithm, learning tasks, computing strategies, and audit rules in the post-processing stage, which ensures the standardization and consistency of the federated learning process and lays the foundation for subsequent privacy protection and efficient training. Constructing a proxy data set and training it to form a proxy model are key steps to protect the privacy of the original data, so that the original data will not be directly exposed during the training process, thereby effectively preventing the risk of data leakage. At the same time, the proxy model can simulate the characteristics of the original data to a certain extent, ensuring the effectiveness and accuracy of the training. Each device participant performs customized optimization based on the learning tasks and computing strategies of the proxy model, which improves the flexibility and adaptability of federated learning. The proxy model is optimized using logical optimization, which can significantly It significantly reduces computational complexity, improves model training speed, realizes efficient federated learning training, reduces network bandwidth consumption, improves data processing efficiency, establishes a public key infrastructure to identify participants, ensures the security of communication data, and establishes a public key infrastructure to assign a unique public and private key pair to each participant, thereby achieving data confidentiality and integrity protection. At the same time, the establishment of a public key infrastructure also supports security audit functions, providing a reliable basis for audits in the post-processing stage, and establishing an audit strategy to audit the data input by each device participant, ensuring data compliance and consistency. Through the audit process, third-party auditors can verify whether the data input by each party complies with the protocol provisions and whether the model update is trained based on real and valid data, which further enhances the transparency and credibility of the federated learning process.

[0058] It should be noted that the agreement covers data preprocessing, model training algorithms, learning tasks, computing strategies, and audit rules for the post-processing stage, ensuring that all participants have a common understanding and agreement on the training process; the proxy model learns the distribution and characteristics of the data during the training process, but does not directly process the original sensitive data; in addition, after the participants reach a federated learning training agreement, they need to establish the calculations that need to be run. This step is not only related to the specific content of the learning task, but also includes the key computing strategies that determine the training process:

[0059] Participants need to clarify the specific learning tasks of federated learning. These tasks may involve various machine learning problems such as classification, regression, and clustering. Different tasks correspond to different model architectures and training objectives. By determining the learning tasks, participants can ensure that all participants are working towards a common goal. This is the basis for collaborative training of federated learning.

[0060] The formulation of computing strategies is the key to ensuring efficient, accurate and secure federated learning. These strategies may involve specific parameter settings for model training, such as learning rate, batch size, etc. These parameters have an important impact on the convergence speed and performance of the model. In addition, the number of iterations is also a factor that needs to be carefully considered. Too many iterations may lead to overfitting, while too few iterations may lead to insufficient model training. Participants need to determine the optimal number of iterations through experiments and adjustments based on specific tasks and data characteristics.

[0061] In addition to the basic parameters of model training, the choice of optimization algorithm is also an important part of the computing strategy. The optimization algorithm determines the way the model parameters are updated, which directly affects the training speed and effect of the model. Common optimization algorithms include gradient descent, momentum method, Adam, etc. Each algorithm has its own characteristics and applicable scenarios. Participants need to choose the most appropriate optimization algorithm based on task requirements and resource constraints.

[0062] In the process of establishing computing strategies, participants also need to consider the issue of data privacy protection. One of the core advantages of federated learning is to conduct collaborative training while keeping data decentralized and private. Therefore, in the formulation of computing strategies, participants need to adopt privacy protection technologies such as secure multi-party computing (MPC) to ensure that the original data of any participant is not leaked during the calculation process. Through carefully designed computing strategies, participants can achieve efficient and accurate federated learning training while protecting data privacy.

[0063] Preferably, constructing a proxy data set according to the protocol and training to form a proxy model specifically includes:

[0064] Obtain original data sets from each device participant;

[0065] Determine the sample space of the original data set and extract features covering the sample space;

[0066] Create a virtual data set through a noise generating function;

[0067] Mix the original dataset with the virtual dataset to form a proxy dataset;

[0068] According to the protocol content, the proxy model is trained using the proxy dataset and loss function.

[0069] In this embodiment, a virtual data set, i.e., noise samples, is introduced and mixed with the original data set to form a proxy data set. This method provides a sufficient amount of data for model training while keeping the original data from being leaked, which makes it difficult for attackers to infer specific information of the original data from the model or the training process, thereby enhancing the privacy protection capability. The proxy data set contains noise samples, which will have a certain impact on the model during the training process, so that the model not only learns to extract features from real data, but also learns to tolerate and identify noise. This training method helps to improve the robustness of the model, making it more resistant to interference from noise and abnormal data in practical applications. The construction process of the proxy data set can be carried out according to actual needs. Adjustment, such as controlling the characteristics of the virtual dataset by changing the parameters of the noise generation function. This flexibility enables the technical solution to be applicable to different application scenarios and datasets, enhancing its adaptability and practicality. Using proxy datasets for model training can be regarded as a form of data enhancement. By introducing more data changes during the training process, the model can learn richer and more comprehensive feature representations, thereby improving the model's generalization ability and prediction accuracy. The technical solution operates strictly in accordance with the content of the agreement. From obtaining the original dataset to training the proxy model, each step follows the process and standards specified in the agreement. This standardization ensures the repeatability and controllability of the training process, and provides strong support for subsequent model evaluation and auditing.

[0070] It should be noted that to determine the sample domain: first, it is necessary to determine the sample space X of the original dataset Dorig and ensure that the sample space Xproxy of the proxy dataset covers the key features of X.

[0071] Construct a virtual dataset: Create a virtual dataset Dvirt through the noise generating function g(μ,σ2) (where μ and σ2 are the mean and variance respectively), whose samples can be expressed as xvirt-g(μ,σ2).

[0072] Mixed data set: The original data set Dorig and the virtual data set Dvirt are mixed according to the mixing ratio α to form a proxy data set Dproxy, that is, Dproxy = αDorig∪(1-α)Dvirt.

[0073] Training the proxy model: Using the proxy dataset Dproxy and loss function To train the proxy model Mproxy, the training process can be expressed as an optimization problem, that is, finding the optimal model parameter θ* so that the average loss on Dproxy is minimized.

[0074] Preferably, each device participant performs customized optimization of learning tasks and strategies based on the learning tasks and computing strategies of the proxy model, specifically including:

[0075] Obtain learning tasks and strategies customized by each device participant;

[0076] Generate DSL code for each device participant based on the learning tasks and strategies customized by each device participant;

[0077] Call the API interface to execute the customized learning tasks and strategies of each device participant, and optimize the learning tasks and computing strategies of the agent model.

[0078] In this embodiment, each participant can flexibly define learning tasks and computing strategies according to their own specific needs and data characteristics. This flexibility ensures that the federated learning solution can be widely applied to different application scenarios and data environments. Through DSL (domain-specific language) and API interface, each participant can convert their own customized learning tasks and strategies into executable code. This personalized optimization strategy helps to improve the pertinence and effectiveness of model training. The optimization of computing strategy covers multiple aspects such as the number of iterations of model training, optimization algorithm, aggregation strategy, etc. Through refined computing strategy design, the efficiency and accuracy of model training can be improved while protecting data privacy. The parties can clearly define the specific learning tasks of federated learning, such as classification and regression, and customize the corresponding model architecture and training objectives according to the task requirements. This customized learning task helps to ensure the pertinence and effectiveness of model training; through the customized optimization strategies of each participant, better synergy and higher training efficiency can be achieved in the collaborative training process of federated learning, which helps to reduce unnecessary computing overhead and accelerate the convergence process of the model; in the process of customizing learning tasks and strategies, by adopting privacy protection technologies such as secure multi-party computing (MPC), it can be ensured that data privacy is effectively protected during the calculation process, which further enhances the privacy protection capability of the federated learning solution.

[0079] It should be noted that DSL provides participants with a unified and structured language to define learning tasks and strategies. This language is not only easy to understand and use, but also can clearly describe all aspects of the task, including data preprocessing, model training algorithm selection, number of iterations, optimization algorithms, etc. Participants can use DSL to write task and strategy descriptions that meet their needs based on their own professional knowledge and data characteristics.

[0080] Through the API, participants can seamlessly integrate custom tasks and strategies written using DSL into the federated learning system. The API provides participants with an interface for interacting with the system, allowing them to easily pass their own tasks and strategies to the system for execution. In this way, participants can not only flexibly adjust and optimize their own tasks and strategies, but also share and collaborate with other participants to complete tasks, thereby jointly promoting the performance improvement of the federated learning system.

[0081] In the process of customizing learning tasks and expressing strategies, the system also provides mechanisms such as logic optimization to accelerate the aggregation process. These mechanisms improve the efficiency and performance of federated learning by optimizing aggregation algorithms, reducing data transmission volume, and utilizing distributed computing resources. At the same time, the establishment of a public key infrastructure (PKI) also provides guarantees for participant identity identification and data transmission security.

[0082] Preferably, the optimizing the proxy model by means of logic optimization specifically includes:

[0083] Adopt gradient compression technology to optimize the aggregation algorithm of proxy model;

[0084] Distributed computing resources are utilized and parallel computing technology is adopted to accelerate the aggregate update of agent model parameters.

[0085] In this embodiment, the gradient compression technology significantly reduces the consumption of network bandwidth by reducing the amount of data of gradient information that needs to be transmitted during the federated learning process, which not only speeds up data transmission, but also enables more efficient model training in a bandwidth-constrained environment; due to the reduction of data transmission delay and overhead, the gradient compression technology makes the update of model parameters faster, thereby improving the training efficiency of the entire federated learning process; during the compression process, the gradient information may be encrypted or transformed, further enhancing the privacy protection of the data during transmission; by utilizing distributed computing resources and assigning computing tasks to multiple nodes for simultaneous processing through parallel computing technology, the aggregation update process of the proxy model parameters can be significantly accelerated, and this parallel processing capability makes it possible to perform efficient training on large-scale data sets; distributed computing resources are fully utilized, avoiding the bottleneck of single node computing power and improving the utilization rate of overall computing resources; with the increase in the number of participants and the expansion of computing resources, the technical solution can maintain an efficient training speed and has good scalability.

[0086] It should be noted that aggregation algorithm optimization: optimize the traditional aggregation algorithm Agg by reducing the amount of data transmission (such as gradient compression) and using parallel computing. trad , and get the optimized aggregation algorithm Agg trad , Gradient compression can use quantization or sparsification techniques.

[0087] Distributed computing resource utilization: In a distributed environment, the optimized aggregation algorithm is used to aggregate and update model parameters, that is, Where N is the number of parties.

[0088] Preferably, the establishment of a public key infrastructure, wherein the public key infrastructure is used to identify the identity of the participant, specifically includes:

[0089] Each participant is assigned a unique public and private key pair for encryption and decryption of communication data;

[0090] The identity of the participants is verified through a digital certificate mechanism to ensure the authenticity of the identities of both communicating parties and the integrity of the data.

[0091] In this embodiment, each participant has a unique public key and private key pair. The public key is used to encrypt data, while the private key is used to decrypt data. This mechanism ensures that only legitimate recipients can decrypt and read communication data, thereby preventing data from being stolen or tampered with by unauthorized third parties during transmission. Through the digital certificate mechanism, the key information such as the participant's public key, identity information, and certificate validity period are issued and verified by a trusted certificate authority (CA). This mechanism ensures that both parties in communication can verify the authenticity of each other's identity and prevent identity forgery or impersonation. The digital certificate not only verifies the identity of the participant, but also ensures the integrity and consistency of the transmitted data through the encryption and decryption process of the public key and private key. Any tampering or destruction of the data can be detected through signature verification, thereby ensuring the reliability and credibility of the data. PKI also supports security auditing functions, and records the complete transmission history of data through technologies such as digital signatures and timestamps, including key information such as the sender, receiver, transmission time and content of the data. This provides an audit basis for third-party auditors, and can trace and hold accountable the behavior of the participants in the post-processing stage to ensure that all parties comply with the agreement.

[0092] It should be noted that the public key generation and distribution: for each participant P i Generate a public-private key pair (pk i ,sk i ) and register the public key in PKI for data encryption.

[0093] Data commitment and signature: Participants calculate data commitment comm i and signature sig i to ensure the integrity and authenticity of the data.

[0094] Audit process: Audit entity A verifies data commitments and signatures to ensure that the data has not been tampered with and flags suspicious parties.

[0095] Preferably, the establishment of an audit strategy, which is used to audit the data input by each device participant, specifically includes:

[0096] Each participant calculates the committed amount of their data and generates a signature using a private key;

[0097] All parties run secure computations encoding the original learning task and check the consistency of the input data;

[0098] Verify the input data of each participant based on the public key infrastructure (PKI) and detect potential malicious participants.

[0099] In this embodiment, each participant calculates the committed amount of its data and generates a signature using a private key. This process ensures that the content and source of the data are traceable and verifiable before it is submitted to the federated learning system. This helps ensure that the data submitted by each party complies with the protocol and enhances the consistency and compliance of the data. Using a public key infrastructure (PKI) to verify the input data of each participant can ensure the integrity and authenticity of the data. Through digital signatures and certificate verification, data can be effectively prevented from being tampered with or forged during transmission, thereby protecting the security of the data. The inspection mechanism in the audit policy can identify inconsistent input data and further detect potential malicious participants through PKI verification. Once data inconsistency or abnormal behavior of participants is found, the system can immediately take measures, such as terminating calculations or reporting to relevant agencies, thereby effectively preventing malicious participants from influencing and damaging the federated learning model; by establishing a transparent and verifiable audit process, all participants can have more trust in the fairness and security of the entire federated learning system, which helps promote cooperation and trust among all participants and improve the overall success rate and efficiency of federated learning projects; the implementation of audit strategies enables the system to respond and recover quickly in the face of malicious attacks or data tampering. By promptly discovering and processing abnormal data or participants, the system can maintain its stability and robustness, ensuring that the federated learning task can proceed smoothly.

[0100] It should be noted that, using the established PKI, Pi has a public key corresponding to each participant in the secure computation, and Pi agrees to use the same unique number qid.

[0101] Pi calculates the commitment of its data, let the commitment be c i , Pi hash commitment h i = hash(c i ), and uses its secret key to generate a signature σ i =sign(qid,h i ), Pi will (c i , σ i )Published to P j≠i·

[0102] All Pis run a secure computation that encodes the original learning task, as well as a preprocessing phase that checks whether Pi's input data is indeed committed to the public commitments received by the parties from Pi. If the check fails, the computation is aborted. Note that at this stage we do not know who cheated, but the parties will not get any results either, because the computation will be aborted before any part of the learning task is executed.

[0103] During the audit, Pi will publish its signed commitment letter to A and the (c j , σ j ), A checks whether all the commitments received from Pj about Pi match, if not, Pi is detected as malicious.

[0104] A runs a two-party secure computation for each Pi separately. Pi inputs the data, and A checks the data against the corresponding commitment. If they match, the audit function continues. If this computation is aborted, Pi will also be detected as malicious. Since the audit is performed in the secure computation, A will not directly see Pi's input data.

[0105] A privacy protection system for federated learning based on proxy datasets, comprising:

[0106] The agreement module is used for multiple device participants to reach an agreement before participating in training;

[0107] A proxy data set construction and model training module, used to construct a proxy data set according to the protocol and train to form a proxy model;

[0108] A setting module, used to set the learning tasks and calculation strategies of the agent model;

[0109] The custom optimization module is used by each device participant to perform custom optimization of learning tasks and strategies based on the learning tasks and computing strategies of the agent model;

[0110] A logic optimization module, used for optimizing the proxy model by using a logic optimization method;

[0111] A public key infrastructure module, used for establishing a public key infrastructure, wherein the public key infrastructure is used for identifying the identity of a participant;

[0112] The audit policy module is used to establish an audit policy, and the audit policy is used to audit the data input by each device participant.

[0113] A computing device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above-mentioned method for privacy protection of federated learning based on a proxy data set are implemented.

[0114] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-mentioned method for privacy protection of federated learning based on a proxy data set.

[0115] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for privacy protection of federated learning based on proxy datasets, characterized in that: Specifically include: Multiple device participants reach an agreement before participating in training; Constructing a proxy data set according to the protocol and training it to form a proxy model; Setting the learning tasks and computing strategies of the agent model; Each device participant performs customized optimization of learning tasks and strategies based on the learning tasks and computing strategies of the agent model; Optimizing the proxy model using a logic optimization method; Establishing a public key infrastructure for identifying participants; An audit policy is established, wherein the audit policy is used to audit the data input by each device participant.

2. According to claim 1, a method for protecting privacy in federated learning based on a proxy dataset is characterized in that: The step of constructing a proxy data set and training a proxy model according to the protocol specifically includes: Obtain original data sets from each device participant; Determine the sample space of the original data set and extract features covering the sample space; Create a virtual data set through a noise generating function; Mix the original dataset with the virtual dataset to form a proxy dataset; According to the protocol content, the proxy model is trained using the proxy dataset and loss function.

3. A method for protecting privacy in federated learning based on a proxy dataset according to claim 2, characterized in that: Each device participant performs customized optimization of learning tasks and strategies based on the learning tasks and computing strategies of the proxy model, specifically including: Obtain learning tasks and strategies customized by each device participant; Generate DSL code for each device participant based on the learning tasks and strategies customized by each device participant; Call the API interface to execute the customized learning tasks and strategies of each device participant, and optimize the learning tasks and computing strategies of the agent model.

4. A method for protecting privacy in federated learning based on a proxy dataset according to claim 3, characterized in that: The optimization of the proxy model by means of logic optimization specifically includes: Adopt gradient compression technology to optimize the aggregation algorithm of proxy model; Distributed computing resources are utilized and parallel computing technology is adopted to accelerate the aggregate update of agent model parameters.

5. A method for protecting privacy in federated learning based on a proxy dataset according to claim 4, characterized in that: The establishment of a public key infrastructure, which is used to identify the identity of the participant, specifically includes: Each participant is assigned a unique public and private key pair for encryption and decryption of communication data; The identity of the participants is verified through a digital certificate mechanism to ensure the authenticity of the identities of both communicating parties and the integrity of the data.

6. A method for protecting privacy in federated learning based on a proxy dataset according to claim 5, characterized in that: The establishment of an audit strategy, which is used to audit the data input by each device participant, specifically includes: Each participant calculates the committed amount of their data and generates a signature using a private key; All parties run secure computations encoding the original learning task and check the consistency of the input data; Verify the input data of each participant based on the public key infrastructure (PKI) and detect potential malicious participants.

7. A privacy protection system for federated learning based on proxy datasets, characterized in that: include: The agreement module is used for multiple device participants to reach an agreement before participating in training; A proxy data set construction and model training module, used to construct a proxy data set according to the protocol and train to form a proxy model; A setting module, used to set the learning tasks and calculation strategies of the agent model; The custom optimization module is used by each device participant to perform custom optimization of learning tasks and strategies based on the learning tasks and computing strategies of the agent model; A logic optimization module, used for optimizing the proxy model by using a logic optimization method; A public key infrastructure module, used for establishing a public key infrastructure, wherein the public key infrastructure is used for identifying the identity of a participant; The audit policy module is used to establish an audit policy, and the audit policy is used to audit the data input by each device participant.

8. A computing device, characterized in that The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the method for privacy protection of federated learning based on a proxy data set described in any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, implements the steps of the method for privacy protection of federated learning based on a proxy data set as described in any one of claims 1 to 6.

Images