Decentralized collaborative learning method and system for privacy protection and computer equipment

Abstract

The invention belongs to the technical field of computers, and relates to a privacy-protected decentralized collaborative learning method and system and computer equipment. The method comprises the following steps: each client calculates a local gradient by using local data and a current local model, generates Gaussian noise and adds the Gaussian noise into the local gradient for disturbance; sending the current local model to a neighbor; each client calculates a corresponding cross gradient by using own data and a local model of a neighbor, injects Gaussian noise into the gradient for disturbance, and returns the disturbed gradient to the corresponding neighbor; each client side calculates the Sharpley values of neighbors and the client side, and determines the weight during aggregation and the local gradient and the cross gradient after aggregation disturbance; updating the current local model according to the aggregated gradient to obtain a latest local model; and iterating for multiple times until a preset learning round number is reached or a loss value reaches a specified threshold value. According to the method, the privacy data of the client is protected, the global model performance is improved, and the convergence of the model is accelerated.

Description

Technical Field

[0001] The present invention belongs to the field of computer technology and relates to a privacy-protected decentralized collaborative learning method, system and computer equipment. Background Art

[0002] Collaborative learning uses the local data and computing power of edge terminal devices to train learning models or parameters, and then shares and integrates model parameter updates through communication between terminal devices, forming a collaborative edge intelligent system. According to the communication topology, collaborative learning can be divided into centralized collaborative learning and decentralized collaborative learning. In centralized collaborative learning, edge devices need to communicate with the central server continuously. When the edge devices are massive or billions of scale, this increases the resource consumption of the central server. In addition, when the central server becomes unreliable, this makes the entire learning process difficult. Therefore, decentralized collaborative learning came into being. Decentralized collaborative learning means that edge devices do not need to rely on any central node, but achieve distributed model training through point-to-point communication and collaboration. In decentralized collaborative learning, due to the different environments and collection preferences of each edge device, the training data between different edge devices are very different in order of magnitude, quantity category and acquired features, thus forming data heterogeneity, that is, the data non-independently identically distributed (Non-IID) phenomenon.

[0003] The non-independent and identically distributed data will affect the ability of decentralized collaborative learning and increase the difficulty of learning. Therefore, the problem of non-independent and identically distributed data has become a research focus of decentralized collaborative learning. Although the existing decentralized learning algorithms avoid direct sharing of original data through local training and gradient exchange, thereby reducing the risk of sensitive information leakage, privacy protection is still an important challenge they face. Studies have shown that malicious attackers can still infer and extract privacy-sensitive information from shared model parameters or gradients, posing a threat to privacy security. Therefore, how to simultaneously solve the problems of data heterogeneity and privacy leakage in decentralized learning and improve the robustness of the global model is a problem to be studied.

[0004] In the patent CN118966384A robust decentralized collaborative learning method and system, computer device and medium, although it is disclosed that the client calculates the cross gradient for the neighbor and directly transmits the calculated cross gradient back to the corresponding neighbor; however, there is a problem of privacy data leakage in this patent. Studies have shown that attackers construct virtual data and minimize the difference between the virtual gradient (the gradient obtained based on the virtual data as input) and the shared real gradient, so that the virtual data is close to the original private training data, thereby analyzing the client's privacy data and causing the leakage of privacy data. In response to this problem, the present invention improves the patent "CN118966384A robust decentralized collaborative learning method and system, computer device and medium", and on this basis, integrates the differential privacy mechanism to protect the client data, and uses the Shapley value to alleviate the negative impact of noise introduction on data availability. Summary of the invention

[0005] In view of the above defects and improvement needs of the prior art, the present invention provides a privacy-preserving decentralized collaborative learning method. The present invention uses the differential privacy mechanism to add Gaussian noise to the local gradient information and cross-gradient information of the client to interfere, thereby effectively protecting the local privacy data of each client. In order to reduce the impact of noise, the present invention uses the Shapley value to fairly evaluate the contribution of the perturbed local gradient and the received perturbed cross-gradient to the global learning objective, thereby improving the robustness of the global model and ensuring the learning performance of the model.

[0006] One of the technical solutions provided by the present invention to solve the technical problem is: A privacy-preserving decentralized collaborative learning method, the method comprising the following steps: (1) Each client uses local data and the current local model to calculate the local gradient; (2) Each client generates Gaussian noise and adds the generated Gaussian noise to the local gradient for perturbation to obtain the perturbed local gradient; (3) According to the communication topology, each client sends the current local model to its neighbors; (4) After each client receives the neighbor’s local model, it uses the local data and the neighbor’s local model to calculate the corresponding cross gradient; (5) Each client generates Gaussian noise for the cross-gradient of each neighbor, adds the Gaussian noise to the cross-gradient for perturbation to protect its own privacy data, and transmits the perturbed cross-gradient back to the corresponding neighbor; (6) Based on the perturbed local gradient and the received perturbed cross gradient, each client uses the validation set to calculate the Shapley value of its neighbors and itself, thereby determining the weight assigned to each client; (7) Each client aggregates the perturbed local gradient and the perturbed cross gradient according to the weights assigned in step (6) to obtain the aggregated gradient, and updates the current local model according to the aggregated gradient to obtain the latest local model; (8) Repeat steps (1) to (7) until the preset number of learning rounds is reached or the loss value reaches the specified threshold, and then stop learning.

[0007] Furthermore, in step (2), the method of performing noise perturbation on the local gradient includes: Clip local gradients: ; in, is the local gradient of client i after clipping in round t; is the local gradient of client i in round t; C is the pre-set clipping threshold, min{} means taking the minimum value of the set, Represents the local gradient The square value of the second norm of ; Get the clipped local gradient After that, Gaussian noise is injected into the clipped local gradient: ; in, is the local gradient after noise perturbation; is a Gaussian distribution, is the noise variance; is the identity matrix of dimension d.

[0008] Furthermore, in step (5), Clip the cross gradients: ; in, is the clipped cross gradient; The client uses local data For Neighbors j The calculated cross gradient; C is the preset clipping threshold; min{} means taking the minimum value of the set, Represents cross gradient The square value of the second norm of ; After obtaining the clipped cross gradient, inject Gaussian noise into the clipped cross gradient: ; in, is the cross gradient after noise perturbation; is a Gaussian distribution, is the noise variance; is the identity matrix of dimension d.

[0009] Furthermore, in step (6), each client uses the verification set to calculate the Shapley value of its neighbors and itself as follows: make represents a set of N clients, represents the set including client i and its neighbors; is the characteristic function of the cooperative game, representing the set The contribution of the clients in to the decentralized collaborative learning task; , Is a collection A random permutation of the clients in ; Among them, the characteristic function v is defined as follows: ; ; in, Represents the validation dataset saved by each client that reflects the global learning objective ; To verify the number of data set samples, data samples ; is the indicator function; Representation Model In the data sample The prediction results on is to aggregate each client in round t of The model obtained, , According to the client of round t j The model obtained by the perturbation gradient contributed by client i; Client The Shapley value in round t Defined as ; in, For collection Length; By normalization, we get Normalized Shapley value in the interval ; Using the normalized Shapley value Metrics Client Contribute in round t.

[0010] Furthermore, in step (6), the weight assigned to each client is determined based on the obtained Shapley value of the client: ; in, For Clients The weight value in round t; is the connection weight corresponding to the communication topology graph between client i and client j; is the normalized Shapley value calculated by client i for client k in round t.

[0011] Furthermore, in step (7), for each client i, the way to aggregate the perturbed local gradient and the perturbed cross-gradient information from neighbors is as follows: ; in, is the cross gradient after noise perturbation received by neighbor j from client i, For Clients The weight value in round t; is the aggregated gradient of client i; Get the aggregated gradient at each client i After that, the local model is updated. The update process includes: First update the momentum: ; in, is the momentum transmitted by neighbor j to client i; for client j, , is the momentum factor, is the momentum vector of client j in the t-1th round; Then update the local model: ; in, Updated local model for client i, is the model of the tth round transmitted by neighbor j to client i, that is, ; is the local model of client j in round t-1, is the learning rate.

[0012] A privacy-preserving decentralized collaborative learning system, comprising: A model sending and receiving module, wherein the client uses the model sending and receiving module to send local model information and perturbed cross gradient information to neighbors, and receive local model information and perturbed cross gradient information sent by neighbors; Model training module, the client uses the model training module to obtain local gradients according to local data and local model training, and add Gaussian noise to the local gradients to obtain perturbed local gradients; obtain cross gradients according to local data and neighbor's local model training, and add Gaussian noise to obtain perturbed cross gradients; A calculation module, each client uses the calculation module, and with the help of a global verification set, calculates the client's Shapley value according to the perturbed local gradient and the perturbed cross gradient transmitted by the neighbor; determines the corresponding weight when the gradient is aggregated according to the client's Shapley value, and then aggregates the perturbed local gradient and cross gradient to obtain the final gradient, and updates the local model according to the obtained final gradient; after multiple iterations, until a preset number of learning rounds is reached or the loss value reaches a specified threshold, the learning is stopped.

[0013] A computer-readable medium stores a computer program, and when the computer program is read and executed by a processor, the decentralized collaborative learning method with privacy protection is implemented.

[0014] A computer device comprises at least one processor and the computer-readable medium. When the computer program in the computer-readable medium is read and executed by the processor, the decentralized collaborative learning method with privacy protection is implemented.

[0015] Compared with the prior art, the present invention has the following beneficial effects: The present invention uses a differential privacy mechanism to add noise to the cross-gradient information, thereby effectively protecting the privacy of client data and reducing the risk of privacy leakage. It also combines the Shapley value technology to reasonably measure the contribution of the perturbed gradient to the global learning task, thereby reducing the impact of noise on the model availability, improving model performance while protecting privacy, enhancing the generalization and reasoning capabilities of the model, and accelerating the convergence of the model.

[0016] Compared with the patent "N118966384A Robust Decentralized Collaborative Learning Method and System, Computer Device and Medium", the present invention allows the client to generate Gaussian noise to interfere with the cross-gradient and local gradient by means of differential privacy mechanism before sending the cross-gradient to its neighbors, so as to prevent attackers from attacking the gradient information. The introduction of noise will reduce the availability of the cross-gradient transmitted by the client, thereby affecting the convergence speed and generalization performance of the global model. To this end, the present invention uses the Shapley value to measure the perturbed cross-gradient and local gradient, thereby reducing the negative impact of the introduction of noise, balancing privacy protection and model performance under the decentralized learning framework, and improving the convergence speed, reasoning and generalization performance of the global model while ensuring the protection of client privacy data. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The overall framework of a privacy-preserving decentralized collaborative learning method disclosed in Example 1 of the present invention; Figure 2 is the average training loss graph of the method in Example 1 of the present invention (MNIST data set); Figure 3 is the average training loss graph of the method in Example 1 of the present invention (CIFAR-10 dataset); Figure 4 This is a test accuracy effect diagram of the method in Example 1 of the present invention (MNIST data set); Figure 5 This is a test accuracy effect diagram of the method in Example 1 of the present invention (CIFAR-10 dataset). DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0019] Example 1: A privacy-preserving decentralized collaborative learning method, such as Figure 1 As shown, the following steps are included: (1) Each client calculates a local gradient using local data and a current local model; specifically, the local data is image data or text data; the local model is a commonly used neural network model, such as a convolutional neural network model, a VGG model, etc.; (2) Each client generates Gaussian noise and adds the generated Gaussian noise to the local gradient for perturbation to obtain the perturbed local gradient; (3) According to the communication topology, each client sends the current local model to its neighbors; (4) After each client receives the neighbor’s local model, it uses the local data and the neighbor’s local model to calculate the corresponding cross gradient; (5) Each client generates Gaussian noise for the cross-gradient of each neighbor, adds the Gaussian noise to the cross-gradient for perturbation to protect its own privacy data, and transmits the perturbed cross-gradient back to the corresponding neighbor; (6) Based on the perturbed local gradient and the received perturbed cross gradient, each client uses the validation set to calculate the Shapley value of its neighbors and itself, thereby determining the weight assigned to each client; (7) Each client aggregates the perturbed local gradient and the perturbed cross gradient according to the weights assigned in step (6) to obtain the aggregated gradient, and updates the current local model according to the aggregated gradient to obtain the latest local model; (8) Repeat steps (1) to (7) until the preset number of learning rounds is reached or the loss value reaches the specified threshold, and then stop learning.

[0020] In step (2) of this embodiment, the method of performing noise perturbation on the local gradient includes: Clip local gradients: ; in, is the local gradient of client i after clipping in round t; is the local gradient of client i in round t; C is the pre-set clipping threshold, min{} means taking the minimum value of the set, Represents the local gradient The square value of the second norm of ; Get the clipped local gradient After that, Gaussian noise is injected into the clipped local gradient: ; in, is the local gradient after noise perturbation; is a Gaussian distribution, is the noise variance; is the identity matrix of dimension d.

[0021] In step (5) of this embodiment, Clip the cross gradients: ; in, is the clipped cross gradient; The client uses local data For Neighbors j The calculated cross gradient; C is the preset clipping threshold; min{} means taking the minimum value of the set, Represents cross gradient The square value of the second norm of ; After obtaining the clipped cross gradient, inject Gaussian noise into the clipped cross gradient: ; in, is the cross gradient after noise perturbation; is a Gaussian distribution, is the noise variance; is the identity matrix of dimension d.

[0022] In step (6) of this embodiment, the method in which each client uses the verification set to calculate the Shapley value of its neighbors and itself is as follows: After receiving the perturbed cross-gradients sent back by the neighbors, each client calculates the neighbor's and its own Shapley value based on the collected perturbed gradient information and the validation set, thereby measuring the client's contribution to the global learning goal and providing appropriate weights for perturbation gradient aggregation. The Shapley value is an important technology in cooperative games. Its core idea is to distribute rewards based on the contribution of each client to the entire cooperation. The Shapley value is calculated as follows: make represents a set of N clients, represents the set including client i and its neighbors; is the characteristic function of the cooperative game, representing the set The contribution of the clients in to the decentralized collaborative learning task; , Is a collection A random permutation of the clients in ; Among them, the characteristic function v is defined as follows: ; ; in, Represents the validation dataset saved by each client that reflects the global learning objective ; To verify the number of data set samples, data samples ; is the indicator function; Representation Model In the data sample The prediction results on is to aggregate each client in round t of The model obtained, , According to the client of round t j The model obtained by the perturbation gradient contributed by client i; Client The Shapley value in round t Defined as ; in, For collection Length; By normalization, we get Normalized Shapley value in the interval ; Using the normalized Shapley value Metrics Client Contribute in round t.

[0023] In step (6), the weight assigned to each client is determined based on the obtained Shapley value of the client: ; in, For Clients The weight value in round t; is the connection weight corresponding to the communication topology graph between client i and client j; is the normalized Shapley value calculated by client i for client k in round t.

[0024] In step (7), for each client i, the way to aggregate the perturbed local gradient and the perturbed cross-gradient information from neighbors is as follows: ; in, is the cross gradient after noise perturbation received by neighbor j from client i, For Clients The weight value in round t; is the aggregated gradient of client i; Get the aggregated gradient at each client i After that, the local model is updated. The update process includes: First update the momentum: ; in, is the momentum transmitted by neighbor j to client i; for client j, , is the momentum factor, is the momentum vector of client j in the t-1th round; Then update the local model: ; in, Updated local model for client i, is the model of the tth round transmitted by neighbor j to client i, that is, ; is the local model of client j in round t-1, is the learning rate.

[0025] Embodiment 2: A privacy-preserving decentralized collaborative learning system, the system comprising: A model sending and receiving module, wherein the client uses the model sending and receiving module to send local model information and perturbed cross gradient information to neighbors, and receive local model information and perturbed cross gradient information sent by neighbors; Model training module, the client uses the model training module to obtain local gradients according to local data and local model training, and add Gaussian noise to the local gradients to obtain perturbed local gradients; obtain cross gradients according to local data and neighbor's local model training, and add Gaussian noise to obtain perturbed cross gradients; A calculation module, each client uses the calculation module, and with the help of a global verification set, calculates the client's Shapley value according to the perturbed local gradient and the perturbed cross gradient transmitted by the neighbor; determines the corresponding weight when the gradient is aggregated according to the client's Shapley value, and then aggregates the perturbed local gradient and cross gradient to obtain the final gradient, and updates the local model according to the obtained final gradient; after multiple iterations, until a preset number of learning rounds is reached or the loss value reaches a specified threshold, the learning is stopped.

[0026] Embodiment 3: A computer-readable medium stores a computer program, and when the computer program is read and executed by a processor, the method described in Embodiment 1 is implemented.

[0027] Embodiment 4: A computer device, characterized in that it comprises at least one processor and the computer-readable medium in Embodiment 1, wherein when the computer program in the computer-readable medium is read and executed by the processor, the method described in Embodiment 1 is implemented. In a specific embodiment, the data sets used in the present invention are MNIST and CIFAR-10 data sets, wherein the MNIST data set is a handwritten digital image data set, including 70,000 handwritten digital images of 10 categories from 0 to 9, of which 60,000 are training images and 10,000 are test images; the CIFAR-10 data set is 60,000 color images of 10 categories: airplanes, cars, birds, cats, deer, dogs, frogs, horses, ships and trucks, of which 50,000 are training images and 10,000 are test images.

[0028] The present invention selects DP-DPSGD algorithm, MUFFLIATO algorithm, DP-CGA algorithm and DP-NET-FLEET algorithm to conduct experimental verification of training set loss value and test accuracy value, wherein DP-DPSGD algorithm is a decentralized learning algorithm combining differential privacy mechanism with distributed SGD method, but this method does not consider the non-independent and identically distributed problem of data between clients; MUFFLIATO algorithm allows each client to add Gaussian noise to local random gradient, and uses multi-round gossip average mechanism to ensure model performance, but this method also does not consider data heterogeneity between clients; DP-NET-FLEET algorithm combines NET-FLEET algorithm with Gaussian mechanism to further ensure data security; in DP-CGA algorithm, each client will use Gaussian noise to perturb it before sharing cross-gradient information with its neighbors, thereby preventing the leakage of local privacy data. Although DP-NET-FLEET algorithm and DP-CGA algorithm consider data heterogeneity between clients, they do not fairly measure the contribution of clients to the global learning goal. In addition, in the experimental setting, the communication topology graph between clients is a fully connected graph, and the total number of clients is 10.

[0029] Centralizing the privacy budget for the MNIST dataset Set to 0.1, the privacy budget in the CIFAR-10 dataset Set to 1. Under this setting, the experimental results of the training set loss value are as follows Figure 2-Figure 3 As shown, Figure 2 This is the result of the training loss value reduction of the MNIST dataset; Figure 3 This is the result of the training loss value reduction of the CIFAR-10 dataset. From Figure - Figure 3 It can be seen that with the increase of the number of training rounds, the training losses of all methods on the MNIST dataset and the CIFAR-10 dataset are decreasing, and the method (PDSL) proposed in the present invention is significantly better than the existing methods, gradually converges with fewer rounds and can reach the minimum value.

[0030] Under different privacy budget settings, that is, the privacy budget in the MNIST dataset Privacy budget in CIFAR-10 dataset is set to 0.08, 0.1 and 0.3 Set to 0.5, 0.7 and 1, the results of testing the exact value are as follows Figure 4 - As shown in 5, where Figure 4 This is the test accuracy result of the MNIST dataset after 180 learning rounds. Figure 5 The test accuracy results of the CIFAR-10 dataset after 200 learning rounds. Figure 4-Figure 5It can be seen that the method (PDSL) proposed in the present invention can have a good test accuracy value, which shows that the method of the present invention can effectively improve the performance of decentralized collaborative learning and improve the robustness of the model while protecting privacy.

[0031] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus dynamic random access memory (Rambus DRAM, referred to as RDRAM), and interface dynamic random access memory (DRDRAM).

[0032] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A privacy-preserving decentralized collaborative learning method, characterized in that: The method comprises the following steps: (1) Each client uses local data and the current local model to calculate the local gradient; (2) Each client generates Gaussian noise and adds the generated Gaussian noise to the local gradient for perturbation to obtain the perturbed local gradient; (3) According to the communication topology, each client sends the current local model to its neighbors; (4) After each client receives the neighbor’s local model, it uses the local data and the neighbor’s local model to calculate the corresponding cross gradient; (5) Each client generates Gaussian noise for the cross-gradient of each neighbor, adds the Gaussian noise to the cross-gradient for perturbation to protect its own privacy data, and transmits the perturbed cross-gradient back to the corresponding neighbor; (6) Based on the perturbed local gradient and the received perturbed cross gradient, each client uses the validation set to calculate the Shapley value of its neighbors and itself, thereby determining the weight assigned to each client; (7) Each client aggregates the perturbed local gradient and the perturbed cross gradient according to the weights assigned in step (6) to obtain the aggregated gradient, and updates the current local model according to the aggregated gradient to obtain the latest local model; (8) Repeat steps (1) to (7) until the preset number of learning rounds is reached or the loss value reaches the specified threshold, and then stop learning.

2. A privacy-preserving decentralized collaborative learning method according to claim 1, characterized in that: In step (2), the method of performing noise perturbation on the local gradient includes: Clip local gradients: ； in, is the local gradient of client i after clipping in round t; is the local gradient of client i in round t; C is the pre-set clipping threshold, min{} means taking the minimum value of the set, Represents the local gradient The square value of the second norm of ; Get the clipped local gradient After that, Gaussian noise is injected into the clipped local gradient: ； in, is the local gradient after noise perturbation; is a Gaussian distribution, is the noise variance; is the identity matrix of dimension d.

3. A privacy-preserving decentralized collaborative learning method according to claim 1, characterized in that: In step (5), Clip the cross gradients: ； in, is the clipped cross gradient; The client uses local data For Neighbors j The calculated cross gradient; C is the preset clipping threshold; min{} means taking the minimum value of the set, Represents cross gradient The square value of the second norm of ; After obtaining the clipped cross gradient, inject Gaussian noise into the clipped cross gradient: ； in, is the cross gradient after noise perturbation; is a Gaussian distribution, is the noise variance; is the identity matrix of dimension d.

4. The privacy-preserving decentralized collaborative learning method according to claim 1, characterized in that: In step (6), each client uses the validation set to calculate the Shapley value of its neighbors and itself as follows: make represents a set of N clients, represents the set including client i and its neighbors; is the characteristic function of the cooperative game, representing the set The contribution of the clients in to the decentralized collaborative learning task; , Is a collection A random permutation of the clients in ; Among them, the characteristic function v is defined as follows: ； ； in, Represents the validation dataset saved by each client that reflects the global learning objective ; To verify the number of data set samples, data samples ; is the indicator function; Representation Model In the data sample The prediction results on is to aggregate each client in round t of The model obtained, , According to the client of round t j The model obtained by the perturbation gradient contributed by client i; Client The Shapley value in round t Defined as ； in, For collection Length; By normalization, we get Normalized Shapley value in the interval ; Using the normalized Shapley value Metrics Client Contribute in round t.

5. The privacy-preserving decentralized collaborative learning method according to claim 4, characterized in that: In step (6), the weight assigned to each client is determined based on the obtained Shapley value of the client: ； in, For Clients The weight value in round t; is the connection weight corresponding to the communication topology graph between client i and client j; is the normalized Shapley value calculated by client i for client k in round t.

6. The privacy-preserving decentralized collaborative learning method according to claim 5, characterized in that: In step (7), for each client i, the way to aggregate the perturbed local gradient and the perturbed cross-gradient information from neighbors is as follows: ； in, is the cross gradient after noise perturbation received by neighbor j from client i, For Clients The weight value in round t; is the aggregated gradient of client i; Get the aggregated gradient at each client i After that, the local model is updated. The update process includes: First update the momentum: ； in, is the momentum transmitted by neighbor j to client i; for client j, , is the momentum factor, is the momentum vector of client j in the t-1th round; Then update the local model: ； in, Updated local model for client i, is the model of the tth round transmitted by neighbor j to client i, that is, ; is the local model of client j in round t-1, is the learning rate.

7. A privacy-preserving decentralized collaborative learning system, characterized in that: The system comprises: A model sending and receiving module, wherein the client uses the model sending and receiving module to send local model information and perturbed cross gradient information to neighbors, and receive local model information and perturbed cross gradient information sent by neighbors; Model training module, the client uses the model training module to obtain local gradients according to local data and local model training, and add Gaussian noise to the local gradients to obtain perturbed local gradients; obtain cross gradients according to local data and neighbor's local model training, and add Gaussian noise to obtain perturbed cross gradients; A calculation module, each client uses the calculation module, and with the help of a global verification set, calculates the client's Shapley value according to the perturbed local gradient and the perturbed cross gradient transmitted by the neighbor; determines the corresponding weight when the gradient is aggregated according to the client's Shapley value, and then aggregates the perturbed local gradient and cross gradient to obtain the final gradient, and updates the local model according to the obtained final gradient; after multiple iterations, until a preset number of learning rounds is reached or the loss value reaches a specified threshold, the learning is stopped.

8. A computer-readable medium, characterized in that The computer-readable medium stores a computer program, and when the computer program is read and executed by a processor, the method according to any one of claims 1 to 6 is implemented.

9. A computer device, characterized in that: The method comprises at least one processor and the computer-readable medium according to claim 8, wherein when the computer program in the computer-readable medium is read and executed by the processor, the method according to any one of claims 1 to 6 is implemented.

Images