Differential privacy federal learning method and system and electronic equipment
By generating and transmitting the optimal contract set in differential privacy federated learning, the client is motivated to inject Gaussian noise and participate in the perturbation of model parameters, the problem of the client not actively participating in federated learning is solved, and the accuracy and privacy protection of the global model are improved.
Patent Information
- Application Number
- CN202411739679.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-29
- Publication Date
- 2025-03-28
AI Technical Summary
The differential privacy federated learning framework does not consider the case where the client does not actively participate in the federated learning process, which affects the accuracy of the global model.
By generating an optimal contract set including multiple optimal contracts, transmitting it to the client, and the client injects Gaussian noise into the privacy protection level in the selected optimal contract, the target model parameters are finally aggregated to obtain the global model.
It improves the initiative of the client to participate in the federated learning process, thereby improving the accuracy of the global model, meeting the client's privacy protection needs and providing reward incentives.
Smart Images
Figure CN119721291A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of Internet of Things security technology, and more specifically, to a differential privacy federated learning method, system and electronic device. Background Art
[0002] In order to cope with the surge in big data from large-scale distributed computing devices (such as mobile phones, drones, and vehicles), distributed machine learning has become increasingly important and indispensable. Currently, the most famous distributed machine learning paradigm is federated learning. In federated learning, multiple clients collaboratively train AI models through a federated server. At the same time, when clients perform local training and exchange AI models, they no longer directly share and publish original data, which brings the privacy protection advantage of preventing data eavesdropping attacks. However, malicious attackers may still be able to infer and extract privacy-sensitive information from public AI models through member inference attacks and model inversion attacks.
[0003] Differential privacy is a promising method to provide standardized privacy guarantees and has been used to prevent data privacy leakage in machine learning. Therefore, differential privacy federated learning has emerged to address the privacy issues at the client level in federated learning. In differential privacy federated learning, each client can perturb its local updates by injecting Gaussian noise, so that better privacy protection can be achieved, that is, a lower degree of privacy leakage. However, the current differential privacy federated learning framework still has many shortcomings. For example: it is usually assumed that clients have a unified privacy attitude in personalized federated systems, ignoring the dynamic impact of client local perturbations on the convergence performance of federated learning algorithms. It is ideally assumed that clients will actively participate in the differential privacy federated learning process, but rarely consider the situation where selfish or inactive clients do not actively participate in the federated learning process due to the additional costs incurred by participating in the federated learning process. Summary of the invention
[0004] The purpose of the present invention is to provide a differential privacy federated learning method, system and electronic device to solve the problem of affecting the accuracy of the global model due to the fact that the differential privacy federated learning framework does not consider the situation where the client does not actively participate in the federated learning process.
[0005] In a first aspect, the present invention provides a differential privacy federated learning method, which is applied to a federated server connected to a plurality of clients for communication, and the method comprises:
[0006] In each round of global communication, an optimal contract set including a plurality of optimal contracts is generated, wherein the optimal contract includes a privacy protection level and a remuneration for the client;
[0007] Transmitting the optimal contract set to each of the clients, and receiving target model parameters returned by each of the clients, wherein the target model parameters are obtained by perturbing the model parameters of the client by injecting Gaussian noise into the model parameters of the client based on the privacy protection level included in the optimal contract selected by the client;
[0008] The target model parameters are aggregated to obtain a global model.
[0009] In a preferred solution, generating an optimal contract set including a plurality of optimal contracts comprises:
[0010] Modeling the interaction process with the client in each round of global communication as a stage privacy protection game;
[0011] Based on the privacy-preserving game at the stages, the interaction with the client in the entire federated learning process is modeled as a multi-stage privacy-preserving game;
[0012] Taking maximizing the target long-term utility function of the multi-stage privacy-preserving game as the goal and satisfying the preset contract theory requirements as the condition, determining the equilibrium solution of the privacy-preserving game in each stage of global communication, wherein the target long-term utility function is the long-term utility function of the federated server;
[0013] The equilibrium solution of the privacy protection game in the stage is used as the optimal contract to generate the optimal contract set.
[0014] In a preferred solution, the privacy protection game at this stage is:
[0015] G t ={{J,S},Γ t ,{U S,t ,{u j,t} j∈J}}
[0016] in, is the optimal contract formulated for the client in the tth round of global communication, the optimal contract Γ t ={ρ j,t ,r j,t} j∈J , ρ j,t is the privacy protection level selected by client j in the tth round of global communication, r j,t is the remuneration paid for the privacy protection level selected by client j in the tth round of global communication; J is the set of clients; S is the federated server; U S,t is the utility function of the federated server in the tth round of global communication; u j,t is the utility function of client j in the tth round of global communication;
[0017] The multi-stage privacy protection game is:
[0018] G={{J,S},{Γ t} t∈T ,{U S,t ,{u j,t} j∈J} t∈T}
[0019] Among them, T is the set of global communications; is the long-term utility function of client j in T global communication rounds; is the target long-term utility function.
[0020] In a preferred solution, the preset contract theory requirements include:
[0021] Contract feasibility, contract optimality, participation fairness, and reward fairness.
[0022] In a preferred solution, the method of maximizing the target long-term utility function of the multi-stage privacy-preserving game and satisfying the preset contract theory requirements is used as a condition to determine the equilibrium solution of the privacy-preserving game in each round of global communication, including:
[0023] By solving the fourth preset formula, determining the equilibrium solution of the privacy protection game at the stage in each round of global communication;
[0024] The fourth preset formula includes:
[0025]
[0026] In a preferred solution, the equilibrium solution of the privacy protection game at the stage includes:
[0027]
[0028] in, is the optimal contract designed for the client j, and the optimal contract meets the preset contract theory requirements:
[0029]
[0030]
[0031] Among them, ω j is the privacy protection requirement of client J; α is the learning rate; ψ is the convergence rate.
[0032] In a preferred solution, before taking the equilibrium solution of the privacy-preserving game at the stage as the optimal contract and generating the optimal contract set, the method further includes:
[0033] Determine whether the equilibrium solution of the privacy-preserving game in the stage satisfies the monotonicity constraint;
[0034] The privacy protection level in the equilibrium solution of the privacy protection game in the stage that does not meet the monotonicity constraint is dynamically adjusted using the "clustering and ironing" algorithm.
[0035] In a second aspect, the present invention also provides another differential privacy federated learning method, which is applied to a client communicating with a federated server, comprising:
[0036] In each round of global communication, obtain model parameters;
[0037] Using zero-concentration differential privacy of Gaussian mechanism, the model parameters are perturbed by injecting Gaussian noise into the model parameters, and the privacy loss is quantified;
[0038] The disturbed model parameters are transmitted to the federated server for aggregation to obtain a global model.
[0039] In a preferred solution, the zero-concentration differential privacy using the Gaussian mechanism perturbs the model parameters by injecting Gaussian noise into the model parameters and quantifies the privacy loss, including:
[0040] Utilizing the zero-concentration differential privacy of the Gaussian mechanism, and according to a first preset formula, perturbing the model parameters by injecting Gaussian noise into the model parameters;
[0041] The first preset formula includes:
[0042]
[0043] in, and are the model parameters of the global communication of the client m in the tth round and the t-1th round respectively; is the Gaussian noise injected by client m in the tth round of global communication; α is the learning rate; q m is the aggregate weight of client m; L m is the privacy loss function of client m; and according to the second preset formula, the privacy loss is quantified;
[0044] The second preset formula includes:
[0045]
[0046] Where E is the expected operator; ▽L and ▽L′ are adjacent model parameters; θ ∈(1,,∞) is the adjustment factor; O is any output corresponding to the model parameter; ρ is the privacy level parameter, and F is the Gaussian mechanism.
[0047] In a preferred solution, after the model parameters are disturbed, the method further includes:
[0048] Based on a gradient descent algorithm, determine the impact of perturbing the model parameters on the performance of the global model.
[0049] In a preferred solution, determining the effect of perturbing the model parameters on the performance of the global model based on a gradient descent algorithm includes:
[0050] According to a third preset formula, based on a gradient descent algorithm, determining the effect of perturbing the model parameters on the performance of the global model;
[0051] The third preset formula includes:
[0052]
[0053] in, x is the dimension of the model parameters, Λ is the gamma function; ι represents the Lipschitz constant; is the model parameter after perturbation in the tth round of global communication; is the undisturbed model parameter in the tth round of global communication; ψ is the convergence rate; μ is the probability that the preset formula is valid.
[0054] In a preferred solution, the differential privacy federated learning method further includes:
[0055] In each round of global communication, a preset optimal contract set is obtained, wherein the preset optimal contract set stores multiple optimal contracts, and the optimal contract includes a privacy protection level and a remuneration;
[0056] Selecting an optimal contract that meets the privacy protection requirements from the preset optimal contract set;
[0057] Based on the privacy protection level included in the selected optimal contract, the model parameters are perturbed by injecting Gaussian noise into the model parameters.
[0058] In a preferred solution, the selecting the optimal contract that meets the privacy protection requirement from the preset optimal contract set includes:
[0059] With the goal of maximizing the utility of all global communications in the federated learning process, an optimal contract that meets the privacy protection requirements is selected from the preset optimal contract set.
[0060] In a preferred solution, the obtaining of model parameters includes:
[0061] Using the local data of the client to perform model training on the global model parameters broadcast by the federal server to obtain updated local model parameters;
[0062] The updated local model parameters are used as the model parameters.
[0063] In a third aspect, the present invention provides a differential privacy federated learning system applied to a federated server connected to a plurality of clients, comprising:
[0064] A generating unit, configured to generate an optimal contract set including a plurality of optimal contracts in each round of global communication, wherein the optimal contract includes a privacy protection level and a remuneration for the client;
[0065] a transceiver unit, configured to transmit the optimal contract set to each of the clients, and receive target model parameters returned by each of the clients, wherein the target model parameters are obtained by perturbing the model parameters of the client by injecting Gaussian noise into the model parameters of the client based on the privacy protection level included in the optimal contract selected by the client;
[0066] The aggregation unit is used to aggregate the target model parameters to obtain a global model.
[0067] In a fourth aspect, the present invention provides a differential privacy federated learning system applied to a client communicating with a federated server, comprising:
[0068] The acquisition unit is used to obtain model parameters in each round of global communication;
[0069] A processing unit, configured to utilize zero-concentration differential privacy of a Gaussian mechanism to perturb the model parameters by injecting Gaussian noise into the model parameters and quantify the privacy loss;
[0070] The transmission unit is used to transmit the disturbed model parameters to the federal server for aggregation to obtain a global model.
[0071] In a fifth aspect, the present invention provides an electronic device, the electronic device comprising:
[0072] processor;
[0073] a memory for storing instructions executable by the processor;
[0074] The processor is used to execute the differential privacy federated learning method provided by the first aspect or the second aspect of the present invention.
[0075] In order to achieve the above-mentioned purpose, the differential privacy federated learning method, system and electronic device provided by the present invention generate an optimal contract set including multiple optimal contracts including the privacy protection level and remuneration of the client in each round of global communication, transmit the optimal contract set to each client, and receive the target model parameters obtained by injecting Gaussian noise into the model parameters of the client to perturb the model parameters of the client based on the privacy protection level included in the optimal contract selected by the client, and aggregate the target model parameters to obtain a global model. Thus, by satisfying the privacy protection requirements of the client and providing remuneration incentives to the client, the initiative of the client to participate in the federated learning process is improved, thereby improving the accuracy of the global model. BRIEF DESCRIPTION OF THE DRAWINGS
[0076] Figure 1 A flowchart of a differential privacy federated learning method provided in an embodiment of the present invention.
[0077] Figure 2 A flowchart of the equilibrium of the privacy protection game in the solution phase provided by an embodiment of the present invention.
[0078] Figure 3 A flowchart of constructing a multi-stage privacy protection game provided in an embodiment of the present invention.
[0079] Figure 4 A specific flow chart of the client provided in the embodiment of the present invention perturbing the model parameters based on the privacy protection level included in the selected optimal contract in each round of global communication and uploading the perturbed model parameters to the federation server.
[0080] Figure 5 A flowchart of another differential privacy federated learning method provided in an embodiment of the present invention.
[0081] Figure 6 A framework diagram of the differential privacy federated learning method provided for applying an embodiment of the present invention.
[0082] Figure 7 A flowchart of a differential privacy federated learning method provided for applying an embodiment of the present invention.
[0083] Figure 8 A structural diagram of a differential privacy federated learning system provided for an embodiment of the present invention.
[0084] Fig. 9 A structural diagram of another differentially private federated learning system provided for an embodiment of the present invention.
[0085] Fig.10 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0086] The present invention is further described in detail below through the accompanying drawings and specific embodiments.
[0087] The differential privacy federated learning method provided by the present invention is executed on an electronic device, which may be an intelligent terminal device such as a laptop, a personal computer, and a tablet computer, or a server on the network side.
[0088] like Figure 1 As shown, the differential privacy federated learning method provided by the embodiment of the present invention is applied to a federated server connected to a plurality of clients for communication, and mainly includes the following steps:
[0089] 110. In each round of global communication, an optimal contract set including multiple optimal contracts is generated.
[0090] Among them, the optimal contract includes the privacy protection level and remuneration for the client.
[0091] 120. Transmit the optimal contract set to each client, and receive the target model parameters returned by each client.
[0092] The target model parameters are obtained by injecting Gaussian noise into the client's model parameters to perturb the client's model parameters based on the privacy protection level included in the optimal contract selected by the client.
[0093] It is understandable that the federated learning process includes multiple rounds of global communication. In each round of global communication, the client uses its own local data to train the global model parameters broadcast by the federated server, obtains updated local model parameters, and then transmits the updated local model parameters to the federated server after perturbing them for aggregation to obtain the global model until the preset maximum number of global iterations is reached or the global model converges.
[0094] In this embodiment, an optimal contract set including multiple optimal contracts is generated in each round of global communication, and then the client selects the optimal contract from the optimal contract set, and the client's model parameters are perturbed by injecting Gaussian noise into the client's model parameters based on the privacy protection level included in the selected optimal contract, so that the client can get corresponding remuneration based on the selected optimal contract and meet its own privacy protection needs. This forms an incentive for the client and improves the initiative of the client to participate in the federated learning process.
[0095] In an optional embodiment, generating an optimal contract set including a plurality of optimal contracts includes:
[0096] The interaction process with the client in each round of global communication is modeled as a stage privacy-preserving game;
[0097] Based on the stage privacy-preserving game, the interaction with the client in the entire federated learning process is modeled as a multi-stage privacy-preserving game;
[0098] With the goal of maximizing the target long-term utility function of the multi-stage privacy-preserving game and the condition of satisfying the preset contract theory requirements, the equilibrium solution of the stage privacy-preserving game in each round of global communication is determined.
[0099] Among them, the target long-term utility function is the long-term utility function of the federated server;
[0100] The equilibrium solution of the stage privacy protection game is taken as the optimal contract to generate the optimal contract set.
[0101] Furthermore, in each round of global communication t, the stage privacy protection game is:
[0102] G t ={{J,S},Γ t ,{U S,t ,{u j,t} j∈J}} (1)
[0103] in, is the optimal contract formulated for the client in the tth round of global communication, the optimal contract Γ t ={ρ j,t ,r j,t} j∈J , ρ j,t is the privacy protection level selected by client j in the tth round of global communication, r j,t is the remuneration paid for the privacy protection level selected by client j in the tth round of global communication; J is the set of clients; S is the federated server; U S,t is the utility function of the federated server in the tth round of global communication; u j,t is the utility function of client j in the tth round of global communication.
[0104] Specifically, it can be seen from formula (1) that the privacy protection game G t It includes the following components:
[0105] Participants: Game G t The participants in are: clients with different privacy protection requirements (i.e., privacy protection preference types) in set J; and the federated server S.
[0106] More specifically, in the game G tIn the t-th round, the goal of the federated server is to design a set of optimal contracts consisting of feasible privacy protection levels and rewards to minimize the long-term cost in the t-th round of global communication. The goal of each client in the t-th round of global communication is to select the optimal contract item to maximize its utility u j,t .
[0107] It should be noted that in the differential privacy federated learning method provided in the embodiment of the present invention, the interaction between the client and the federated server in the entire federated learning process is modeled as a multi-stage privacy protection game. It can be understood that the multi-stage privacy protection game includes a series of stage privacy protection games corresponding to each global communication round to motivate the client to continuously contribute data and computing resources throughout the federated learning process.
[0108] Therefore, after modeling the interaction process with the client in each round of global communication as a stage privacy protection game, the interaction with the client in the entire federated learning process can be modeled as a multi-stage privacy protection game based on the stage privacy protection game:
[0109] G={{J,S},{Γ t} t∈T ,{U S,t ,{u j,t} j∈J} t∈T} (2)
[0110] Among them, T is the set of global communications; is the long-term utility function of client j in T global communication rounds; is the target long-term utility function.
[0111] Specifically, it can be seen from formula (2) that the multi-stage privacy-preserving game consists of a series of stage privacy-preserving games corresponding to T global communication rounds, including the following components:
[0112] Participants: The participants in game G are: clients with different types of privacy preferences in set J, and the federated server S.
[0113] Furthermore, the federation server aims to design a series of contract menus To minimize the long-term cost in T global communication rounds. The goal of each client is to select the optimal contract item from the contract menu to maximize its long-term utility in T global communication rounds.
[0114] Furthermore, in some optional embodiments, without loss of generality, in order to describe the heterogeneity of the clients, the clients may be arranged in ascending order according to the degree of importance they place on privacy protection, expressed as ω1<…<ω j<…<ω J , where ω j is the degree to which client j values privacy, that is, privacy preference or privacy protection requirement, which represents the degree to which the client values privacy leakage. Based on the additivity of ρ-zero centralized differential privacy and the long-term remuneration paid by the federated server, the long-term utility of client j in the entire federated learning process can be defined as:
[0115]
[0116] Based on the dynamic impact of client perturbations on the convergence performance of differentially private federated learning and the remuneration paid to all clients participating in federated learning, the long-term utility of the federated server can be defined as:
[0117]
[0118] in, is the number of clients j.
[0119] It is understandable that if we want to improve the initiative of clients to participate in the federated learning process through the optimal contract, we need to ensure that the optimal contract can mobilize the initiative of clients to choose the optimal contract. Therefore, it is necessary to ensure that the generated optimal contract meets the requirements of feasibility and fairness.
[0120] Based on this, in an optional embodiment, the contract theory requirements are preset, including:
[0121] Contract feasibility, contract optimality, participation fairness, and reward fairness.
[0122] Specifically, in each round of global communication, if each type of client participating in the federated learning process can obtain maximum and non-negative utility when honestly choosing the optimal contract suitable for its specific type, then the contract is considered feasible. In each round of global communication, for a set of feasible contracts, when the contract minimizes the cost of the federated server, it is considered to be the optimal contract, that is, it satisfies the contract optimality. When the client abides by the contract truthfully, participation fairness is satisfied, that is, the client has no motivation to exit the federated learning process and falsely report the type to obtain a higher reward. At the same time, if clients who choose a lower privacy level in the federated learning process can obtain a lower reward, and clients who do not participate in the federated learning process have no reward, then reward fairness is satisfied.
[0123] More specifically, if a contract satisfies both participation fairness and reward fairness, then it achieves contractual fairness.
[0124] Furthermore, after determining the preset contract theory and the purpose of the multi-stage privacy protection game, we can take maximizing the long-term utility function of the federal server of the multi-stage privacy protection game as the goal, and determine the equilibrium solution of the stage privacy protection game in each round of global communication on the condition of meeting the requirements of the preset contract theory. We can also use the equilibrium solution of the stage privacy protection game as the optimal contract to generate the optimal contract set.
[0125] That is, in an optional embodiment, by solving the fourth preset formula, an equilibrium solution of the stage privacy protection game in each round of global communication is determined;
[0126] Among them, the fourth preset formula includes:
[0127]
[0128] Specifically, in each stage of the privacy protection game, the federated server designs an optimal contract set including privacy protection levels and rewards for clients with different privacy protection requirements to solve the game equilibrium of each stage of the game, thereby motivating the client to provide high-quality local model updates in each round of global communication. Specifically:
[0129] The optimal contract design problem is to maximize the long-term utility function of the federated server while satisfying contract feasibility, contract optimality, and contract fairness; therefore, the privacy protection game G t The optimization problem can be defined as formula (5).
[0130] At the same time, given the information asymmetry between the client and the federated server, the contract theory can be used to solve the privacy protection game G according to the formulated long-term utility function of the participants. t The privacy protection game G t The equilibrium strategy is to design the optimal contract for all client types, that is, At the same time, the feasibility and fairness of the contract are ensured. Therefore, the equilibrium of the multi-stage privacy protection game G can be defined as
[0131] More specifically, as shown in formula (5), the first two constraints are contract feasibility constraints, and the last two constraints ensure that the privacy protection level of the client and the remuneration provided by the federated server to the client are non-negative.
[0132] It can be understood that there is J in formula (5) Problem 1 2 The constraints make it impossible to solve directly. Figure 2 As shown, when solving it, it can be simplified into a set of equivalent constraints, thus obtaining the unconstrained optimal contract design problem: that is:
[0133]
[0134] By solving problem 2, we can get the stage privacy protection game G t The equilibrium is expressed as:
[0135] The optimal contract privacy protection level for the unconstrained optimal contract design problem is: The optimal contract reward for the unconstrained optimal contract design problem is: Based on the stage privacy protection game G t The equilibrium solution of the multi-privacy protection game G is theoretically derived as a set of optimal contract privacy protection level-remuneration pairs corresponding to each round of global communication, that is,
[0136] In an optional embodiment, the equilibrium solution of the stage privacy protection game is used as the optimal contract, and before the optimal contract set is generated, the following is also included:
[0137] Determine whether the equilibrium solution of the privacy protection game in the stage satisfies the monotonicity constraint;
[0138] The privacy protection level in the equilibrium solution of the stage privacy protection game that does not meet the monotonicity constraint is dynamically adjusted using the "clustering and ironing" algorithm.
[0139] Specifically, the federated server sorts the optimal contract privacy protection levels of all clients according to their size. When there is an optimal contract privacy protection level sequence that violates the monotonicity constraint, that is, The optimal contract privacy protection level sequence of the client is adjusted by using the dynamic "clustering and ironing" algorithm, so as to obtain the optimal contract privacy protection level sequence that satisfies the monotonicity constraint, that is, the artificial setting in Subsequently, the optimal contract reward, i.e., remuneration, is calculated based on the adjusted optimal contract data size.
[0140] 130. Aggregate the target model parameters to obtain the global model.
[0141] In this embodiment, Figure 3 As shown in the figure, through measures such as contract menu design, long-term utility construction, and game equilibrium definition, the federated server generates an optimal contract set including the optimal contract that specifies the privacy protection level and remuneration, so that the client can maximize its utility by selecting the optimal contract project, that is, in each round of global communication, the model parameters are perturbed based on the privacy protection level specified by the selected optimal contract and the perturbed model parameters are uploaded to the federated server. Repeat the above steps until the maximum number of global iterations is reached or the model converges.
[0142] In each round of global communication, the federated server aggregates the perturbed local model parameters uploaded by the participating clients based on the optimal privacy protection level selected by the participating clients, expressed as:
[0143]
[0144] in, are the local model parameters after perturbation by client j, i.e., the target model parameters.
[0145] Furthermore, in each round of global communication, the client injects Gaussian noise into the client's model parameters to perturb the client's model parameters based on the privacy protection level included in the selected optimal contract, and uploads the perturbed model parameters to the federated server. Figure 4 As shown, in each round of global communication, after selecting the optimal contract that meets its own privacy protection requirements, client j uses its own local data to train the global model parameters broadcast by the federated server to obtain the updated local model parameters:
[0146]
[0147] Among them, d j is the data size used by client j for model training.
[0148] Subsequently, client j uses zero-centralized differential privacy based on the Gaussian mechanism to perturb the updated local model parameters according to the privacy protection level included in the selected optimal contract:
[0149]
[0150] The local model parameters after disturbance are uploaded to the federated server for model aggregation.
[0151] like Figure 5 As shown, the embodiment of the present invention also provides a differential privacy federated learning method, which is applied to a client communicating with a federated server, and mainly includes the following steps:
[0152] 210. In each round of global communication, obtain model parameters.
[0153] In an optional embodiment, obtaining model parameters includes:
[0154] Use the client's local data to train the global model parameters broadcast by the federated server to obtain updated local model parameters;
[0155] The updated local model parameters are used as model parameters.
[0156] In this embodiment, the client uses the client's local data to perform model training on the global model parameters broadcast by the federation server based on the above formula (8) to obtain updated local model parameters.
[0157] 220. Using the zero-concentration differential privacy of the Gaussian mechanism, the model parameters are perturbed by injecting Gaussian noise into the model parameters and the privacy loss is quantified.
[0158] In this embodiment, by using the zero-concentration differential privacy of the Gaussian mechanism, Gaussian noise is injected into the model parameters so that the client can interfere with its local update, thereby achieving better privacy protection, that is, ensuring a lower degree of privacy leakage, that is, improving the confidentiality of the client's private information. The privacy of users in federated learning can be quantitatively analyzed through privacy loss quantification, so as to accurately pay the corresponding incentives.
[0159] Specifically, in an optional embodiment, the zero-concentration differential privacy of the Gaussian mechanism is used to perturb the model parameters by injecting Gaussian noise into the model parameters and quantifying the privacy loss, including:
[0160] Using the zero-concentration differential privacy of the Gaussian mechanism, the model parameters are perturbed by injecting Gaussian noise into the model parameters according to the first preset formula;
[0161] The first preset formula includes:
[0162]
[0163] in, and are the model parameters of the global communication of client m in the tth round and the t-1th round respectively; is the Gaussian noise injected by client m in the tth round of global communication; α is the learning rate; q m is the aggregate weight of client m; L m is the privacy loss function of client m.
[0164] According to the second preset formula, the privacy loss is quantified;
[0165] The second preset formula includes:
[0166]
[0167] Where E is the expected operator; ▽L and ▽L′ are adjacent model parameters; θ ∈ (1,,∞) is the adjustment factor; O is any output corresponding to the model parameter; ρ is the privacy level parameter, and F is the Gaussian mechanism.
[0168] Specifically, if we let σ 2is the variance of Gaussian noise, then a random scheme {F:R→R|F(▽L)=▽L+N(0,σ 2 )} is considered to be ρ-zero centralized privacy. Furthermore, the larger ρ is, the weaker the privacy protection is, because it makes it easier for attackers to distinguish ▽L and ▽L′. Therefore, ρ is regarded as the privacy protection level of the client. By using the Gaussian mechanism F, each client injects a zero-mean and variance σ before transmitting its updated local model parameters to the federated server. 2 Gaussian noise.
[0169] In an optional embodiment, after the model parameters are disturbed, the method further includes:
[0170] Based on the gradient descent algorithm, determine the impact of perturbing the model parameters on the performance of the global model.
[0171] It should be noted that the following lemma shows that gradient descent exhibits a corrective effect in model training:
[0172] Lemma 1: When For L g - Lipschitz smooth, and have:
[0173]
[0174] in, and They respectively represent the differentially private federated learning model obtained by injecting Gaussian noise into the local model parameters in the tth round of global communication and the baseline federated learning model obtained without injecting Gaussian noise; α is the learning rate; ψ is the convergence rate.
[0175] It can be seen from formula (12) that gradient descent can correct the model divergence caused by local perturbations in early iterations, that is, the noise injected in early iterations has more opportunities to be corrected, so their impact on the convergence performance (the impact on the performance of the global model) is weaker than the noise injected in later iterations.
[0176] Further, in an optional embodiment, based on a gradient descent algorithm, determining the impact of perturbing the model parameters on the performance of the global model includes:
[0177] According to a third preset formula, based on a gradient descent algorithm, determining the effect of perturbing the model parameters on the performance of the global model;
[0178] The third preset formula includes:
[0179]
[0180] in, Λ is the gamma function; ι is the Lipschitz constant.
[0181] In this embodiment, based on the correction effect of the gradient descent algorithm, that is, when For L g - Lipschitz smooth, and When , for the differential privacy federated learning model and the benchmark federated learning model with the same initial model parameters, that is, The relationship between the convergence performance of the global model and the client's perturbation is as shown in formula (13), which further proves the quantitative relationship between the convergence performance of differential privacy federated learning and the client's perturbation, and thus provides a basis for modeling utility functions.
[0182] In an optional embodiment, the differential privacy federated learning method further includes:
[0183] In each round of global communication, the preset optimal contract set is obtained.
[0184] Among them, the preset optimal contract set contains multiple optimal contracts, which include privacy protection level and remuneration.
[0185] The optimal contract that meets the privacy protection requirements is selected from the preset optimal contract set;
[0186] Based on the privacy protection level included in the selected optimal contract, the model parameters are perturbed by injecting Gaussian noise into the model parameters.
[0187] In an optional embodiment, an optimal contract that meets the privacy protection requirement is selected from a preset optimal contract set, including:
[0188] With the goal of maximizing the utility of all global communications in the federated learning process, the optimal contract that meets the privacy protection requirements is selected from the preset optimal contract set.
[0189] 230. The disturbed model parameters are transmitted to the federated server for aggregation to obtain a global model.
[0190] Based on the content of the above embodiment, the framework structure of the differential privacy federated learning method provided by the embodiment of the present invention is as follows: Figure 6 As shown, the overall process is as follows Figure 7 shown.
[0191] The differential privacy federated learning method provided in this embodiment uses a federated learning framework with zero-centralized differential privacy based on the Gaussian mechanism to enable each client connected to the federated server to inject Gaussian noise into its model parameters before transmitting them to the federated server to ensure the confidentiality of the client's private information, and at the same time quantify the client's cumulative privacy loss in the entire federated learning process. Through the correction utility based on the gradient descent algorithm, the influence of the noise injected by the personalized client in the local model parameters on the convergence of the federated learning algorithm, that is, the performance of the global model, is quantified, and the dynamic influence of local perturbations on the performance of the global model in different communication rounds is revealed. By modeling the interaction between the client and the federated server in the entire federated learning process as a multi-stage privacy protection game, which contains a series of stage privacy protection games corresponding to each global communication round, the client is encouraged to continuously contribute data and computing resources in the entire federated learning process. By using contract theory to design an optimal contract set containing privacy protection levels and remuneration for clients with different privacy protection requirements, the client is encouraged to provide high-quality models in each round of global communication, while satisfying contract optimality, contract feasibility, individual rationality, incentive compatibility and contract fairness. By enabling the federated server to aggregate the model parameters uploaded by the client based on the privacy protection level included in the optimal contract selected by the client to generate new global model parameters, comprehensive consideration of the client's personalized privacy protection needs is achieved.
[0192] The differential privacy federated learning system provided in an embodiment of the present invention is introduced below. The differential privacy federated learning system described below can be considered as a module architecture for implementing the differential privacy federated learning method provided in an embodiment of the present invention; the content described below can be cross-referenced with the above.
[0193] Optional, see Figure 8 , Figure 8 : is a structural block diagram of a differential privacy federated learning system applied to a federated server connected to a plurality of clients in communication with each other, provided by an embodiment of the present invention. The system may include:
[0194] A generating unit 10, configured to generate an optimal contract set including a plurality of optimal contracts in each round of global communication, wherein the optimal contract includes a privacy protection level and a remuneration for the client;
[0195] The transceiver unit 20 is used to transmit the optimal contract set to each client and receive the target model parameters returned by each client, where the target model parameters are obtained by injecting Gaussian noise into the model parameters of the client to perturb the model parameters of the client based on the privacy protection level included in the selected optimal contract;
[0196] The aggregation unit 30 is used to aggregate the target model parameters to obtain a global model.
[0197] Optionally, the generating unit 10 is specifically configured to:
[0198] The interaction process with the client in each round of global communication is modeled as a stage privacy-preserving game;
[0199] Based on the stage privacy-preserving game, the interaction with the client in the entire federated learning process is modeled as a multi-stage privacy-preserving game;
[0200] Taking maximizing the target long-term utility function of the multi-stage privacy protection game as the goal and satisfying the preset contract theory requirements as the condition, the equilibrium solution of the stage privacy protection game in each round of global communication is determined. The target long-term utility function is the long-term utility function of the federated server.
[0201] The equilibrium solution of the stage privacy protection game is taken as the optimal contract to generate the optimal contract set.
[0202] Optionally, pre-set contract theory requirements, including:
[0203] Contract feasibility, contract optimality, participation fairness, and reward fairness.
[0204] Optionally, the generating unit 10 is more specifically used to determine an equilibrium solution of the stage privacy protection game in each round of global communication by solving a fourth preset formula.
[0205] Optionally, the generating unit 10 is further configured to:
[0206] Determine whether the equilibrium solution of the privacy protection game in the stage satisfies the monotonicity constraint;
[0207] The privacy protection level in the equilibrium solution of the stage privacy protection game that does not meet the monotonicity constraint is dynamically adjusted using the "clustering and ironing" algorithm.
[0208] Optional, see Fig. 9 , Fig. 9 : is a structural block diagram of a differential privacy federated learning system applied to a client communicating with a federated server provided by an embodiment of the present invention. The system may include:
[0209] A collection unit 40, used to obtain model parameters in each round of global communication;
[0210] A processing unit 50 is used to utilize the zero-concentration differential privacy of the Gaussian mechanism to perturb the model parameters by injecting Gaussian noise into the model parameters and quantify the privacy loss;
[0211] The transmission unit 60 is used to transmit the disturbed model parameters to the federation server for aggregation to obtain a global model.
[0212] Optionally, the processing unit 50 is specifically configured to:
[0213] Utilizing the zero-concentration differential privacy of the Gaussian mechanism, according to the first preset formula, the model parameters are perturbed by injecting Gaussian noise into the model parameters; according to the second preset formula, the privacy loss is quantified.
[0214] Optionally, the processing unit 50 is further configured to:
[0215] Based on the gradient descent algorithm, determine the impact of perturbing the model parameters on the performance of the global model.
[0216] Optionally, the processing unit 50 is more specifically configured to:
[0217] According to the third preset formula, based on the gradient descent algorithm, the impact of perturbing the model parameters on the performance of the global model is determined.
[0218] Optionally, the processing unit 50 is specifically configured to:
[0219] In each round of global communication, a preset optimal contract set is obtained. The preset optimal contract set contains multiple optimal contracts, and the optimal contract includes the privacy protection level and remuneration.
[0220] The optimal contract that meets the privacy protection requirements is selected from the preset optimal contract set;
[0221] Based on the privacy protection level included in the selected optimal contract, the model parameters are perturbed by injecting Gaussian noise into the model parameters.
[0222] Optionally, the processing unit 50 is more specifically configured to:
[0223] With the goal of maximizing the utility of all global communications in the federated learning process, the optimal contract that meets the privacy protection requirements is selected from the preset optimal contract set.
[0224] Optionally, the collection unit 40 is specifically used for:
[0225] Use the client's local data to train the global model parameters broadcast by the federated server to obtain updated local model parameters;
[0226] The updated local model parameters are used as model parameters.
[0227] Below, reference Fig.10 To describe the electronic device provided by the embodiment of the present application, the electronic device provided by the embodiment may include: at least one processor 100, at least one communication interface 200, at least one memory 300 and at least one communication bus 400;
[0228] In the embodiment of the present invention, the number of the processor 100, the communication interface 200, the memory 300, and the communication bus 400 is at least one, and the processor 100, the communication interface 200, and the memory 300 communicate with each other through the communication bus 400; obviously, Fig.10 The communication connections shown for the processor 100, the communication interface 200, the memory 300, and the communication bus 400 are merely optional;
[0229] Optionally, the communication interface 200 may be an interface of a communication module, such as an interface of a GSM module; the processor 100 may be a central processing unit CPU, or an application specific integrated circuit ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present invention.
[0230] The memory 300 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.
[0231] The processor 100 is specifically used to execute the application program in the memory to implement the steps of the above-mentioned crane hoisting wire rope installation control method.
[0232] The above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the invention.
Claims
1. A differential privacy federated learning method, applied to a federated server connected to a plurality of clients, characterized in that: The method comprises: In each round of global communication, an optimal contract set including a plurality of optimal contracts is generated, wherein the optimal contract includes a privacy protection level and a remuneration for the client; Transmitting the optimal contract set to each of the clients, and receiving target model parameters returned by each of the clients, wherein the target model parameters are obtained by injecting Gaussian noise into the model parameters of the client to perturb the model parameters based on the privacy protection level included in the optimal contract selected by the client; The target model parameters are aggregated to obtain a global model.
2. The differential privacy federated learning method according to claim 1, characterized in that: The generating of an optimal contract set including a plurality of optimal contracts comprises: Modeling the interaction process with the client in each round of global communication as a stage privacy protection game; Based on the privacy-preserving game at the stages, the interaction with the client in the entire federated learning process is modeled as a multi-stage privacy-preserving game; Taking maximizing the target long-term utility function of the multi-stage privacy-preserving game as the goal and satisfying the preset contract theory requirements as the condition, determining the equilibrium solution of the privacy-preserving game in each stage of global communication, wherein the target long-term utility function is the long-term utility function of the federated server; The equilibrium solution of the privacy protection game in the stage is used as the optimal contract to generate the optimal contract set.
3. The differential privacy federated learning method according to claim 2, characterized in that: The privacy protection game in the above stage is: MR t ={{J,S},Γ t ,{IN S,t ,{in j,t } j∈J }} in, is the optimal contract formulated for the client in the tth round of global communication, the optimal contract Γ t ={ρ j,t ,r j,t } j∈J , ρ j,t is the privacy protection level selected by client j in the tth round of global communication, r j,t is the remuneration paid for the privacy protection level selected by client j in the tth round of global communication; J is the set of clients; S is the federated server; U S,t is the utility function of the federated server in the tth round of global communication; u j,t is the utility function of client j in the tth round of global communication; The multi-stage privacy protection game is: G={{J,S},{Γ t } t∈T ,{IN S,t ,{in j,t } j∈J } t∈T } Among them, T is the set of global communications; is the long-term utility function of client j in T global communication rounds; is the target long-term utility function.
4. The differential privacy federated learning method according to claim 3, characterized in that: The theoretical requirements of the preset contract include: Contract feasibility, contract optimality, participation fairness, and reward fairness.
5. The differential privacy federated learning method according to claim 4, characterized in that: The method aims to maximize the target long-term utility function of the multi-stage privacy protection game and satisfies the preset contract theory requirements, and determines the equilibrium solution of the privacy protection game in each round of global communication, including: By solving the fourth preset formula, determining the equilibrium solution of the privacy protection game at the stage in each round of global communication; The fourth preset formula includes:
6. The differential privacy federated learning method according to claim 5, characterized in that: The equilibrium solution of the privacy protection game in the above stage includes: in, is the optimal contract designed for the client j, and the optimal contract meets the preset contract theory requirements: Among them, ω j is the privacy protection requirement of client J; α is the learning rate; ψ is the convergence rate.
7. The differential privacy federated learning method according to claim 2, characterized in that: Before taking the equilibrium solution of the privacy-preserving game at the stage as the optimal contract and generating the optimal contract set, the method further includes: Determine whether the equilibrium solution of the privacy-preserving game in the stage satisfies the monotonicity constraint; The privacy protection level in the equilibrium solution of the privacy protection game in the stage that does not meet the monotonicity constraint is dynamically adjusted using the "clustering and ironing" algorithm.
8. A differentially private federated learning method, applied to a client communicating with a federated server, characterized in that: The method comprises: In each round of global communication, obtain model parameters; Using zero-concentration differential privacy of Gaussian mechanism, the model parameters are perturbed by injecting Gaussian noise into the model parameters, and the privacy loss is quantified; The disturbed model parameters are transmitted to the federated server for aggregation to obtain a global model.
9. The differential privacy federated learning method according to claim 8, characterized in that: The zero-concentration differential privacy using the Gaussian mechanism perturbs the model parameters by injecting Gaussian noise into the model parameters and quantifies the privacy loss, including: Utilizing the zero-concentration differential privacy of the Gaussian mechanism, and according to a first preset formula, perturbing the model parameters by injecting Gaussian noise into the model parameters; The first preset formula includes: Among them, θ t and θ t-1 are the model parameters of the global communication of the client m in the tth round and the t-1th round respectively; is the Gaussian noise injected by client m in the tth round of global communication; α is the learning rate; q m is the aggregate weight of client m; L m is the privacy loss function of client m; quantifying the privacy loss according to a second preset formula; The second preset formula includes: Where, E is the expectation operator; and are adjacent model parameters; θ∈(1,,∞) is the adjustment factor; O is any output corresponding to the model parameter; ρ is the privacy level parameter, and F is the Gaussian mechanism.
10. The differential privacy federated learning method according to claim 8, characterized in that: After the model parameters are disturbed, the method further includes: Based on a gradient descent algorithm, determine the impact of perturbing the model parameters on the performance of the global model.
11. The differential privacy federated learning method according to claim 10, characterized in that: The determining, based on the gradient descent algorithm, the effect of perturbing the model parameters on the performance of the global model comprises: According to a third preset formula, based on a gradient descent algorithm, determining the effect of perturbing the model parameters on the performance of the global model; The third preset formula includes: in, x is the dimension of the model parameters, Λ is the gamma function; ι is the Lipschitz constant; θ t is the model parameter after perturbation in the tth round of global communication; is the undisturbed model parameter in the tth round of global communication; ψ is the convergence rate; μ is the probability that the preset formula is valid.
12. The differential privacy federated learning method according to claim 8, characterized in that: Also includes: In each round of global communication, a preset optimal contract set is obtained, wherein the preset optimal contract set stores multiple optimal contracts, and the optimal contract includes a privacy protection level and a remuneration; Selecting an optimal contract that meets the privacy protection requirements from the preset optimal contract set; Based on the privacy protection level included in the selected optimal contract, the model parameters are perturbed by injecting Gaussian noise into the model parameters.
13. The differential privacy federated learning method according to claim 12, characterized in that: The selecting the optimal contract that meets the privacy protection requirement from the preset optimal contract set includes: With the goal of maximizing the utility of all global communications in the federated learning process, an optimal contract that meets the privacy protection requirements is selected from the preset optimal contract set.
14. The differential privacy federated learning method according to claim 8, characterized in that: The obtaining of model parameters comprises: Using the local data of the client to perform model training on the global model parameters broadcast by the federal server to obtain updated local model parameters; The updated local model parameters are used as the model parameters.
15. A differential privacy federated learning system applied to a federated server connected to a plurality of clients, characterized in that: include: A generating unit, configured to generate an optimal contract set including a plurality of optimal contracts in each round of global communication, wherein the optimal contract includes a privacy protection level and a remuneration for the client; a transceiver unit, configured to transmit the optimal contract set to each of the clients, and receive target model parameters returned by each of the clients, wherein the target model parameters are obtained by injecting Gaussian noise into the model parameters of the client to perturb the model parameters based on the privacy protection level included in the optimal contract selected by the client; The aggregation unit is used to aggregate the target model parameters to obtain a global model.
16. A differential privacy federated learning system applied to a client communicating with a federated server, characterized in that: include: The acquisition unit is used to obtain model parameters in each round of global communication; A processing unit, configured to utilize zero-concentration differential privacy of a Gaussian mechanism to perturb the model parameters by injecting Gaussian noise into the model parameters and quantify the privacy loss; The transmission unit is used to transmit the disturbed model parameters to the federal server for aggregation to obtain a global model.
17. An electronic device, characterized in that: The electronic device comprises: processor; a memory for storing instructions executable by the processor; The processor is used to execute the differential privacy federated learning method described in any one of claims 1 to 7 or any one of claims 8 to 14.