Federal learning Byzantine attack defense method based on gradient clustering and distance filtering
Through the method based on gradient clustering and distance filtering, the client is screened and weights are dynamically allocated, which solves the problem of insufficient adaptability in the face of different attack modes, reduces the overhead of information transmission, and improves the federated learning system Strongness and recognition accuracy.
Patent Information
- Application Number
- CN202410992188.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-23
- Publication Date
- 2024-11-15
AI Technical Summary
The existing Byzantine robust aggregation method has limited adaptability when facing different types of attack modes, difficult to weigh dynamic weight factors, and adds additional information transmission overhead, which violates the goal of federated learning to reduce traffic.
Using the federated learning Byzantine attack defense method based on gradient clustering and distance filtering, by calculating the instantaneous gradient and cumulative gradient of the client, using the DBSCAN clustering algorithm, calculate the distance correlation and Wasserstein distance, filter the client, and dynamically allocate the weights. Reduce unnecessary information transmission.
The defense effect of different types of Byzantine attacks is achieved, the system's robustness and adaptability is improved, communication overhead is reduced, and the recognition accuracy and stability is significantly improved under medium and low proportion attacks.
Smart Images

Figure CN118965360A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of artificial intelligence and network security, and mainly relates to a federated learning Byzantine attack defense method based on gradient clustering and distance filtering. Background Art
[0002] In federated learning, Byzantine attacks are a potential threat that may lead to malicious tampering of model parameters, thus affecting the performance and credibility of the entire federated learning system. Byzantine attacks are usually initiated by some malicious Byzantine nodes, which may deliberately provide false updates or tamper with real model parameters to disrupt the model's convergence process or cause model performance degradation. Although the existing Byzantine robustness aggregation method technical solution has improved the resistance of the federated learning system to Byzantine attacks to a certain extent, it still has some shortcomings:
[0003] (1) Inability to adapt to other types of attack patterns: Technical solutions for Byzantine attacks are usually designed based on a certain type of attack, but have limited adaptability to other types of attack patterns, requiring continuous upgrading of defense strategies.
[0004] (2) It is difficult to weigh the uncertainty of different factors: In terms of dynamic weight allocation, how to determine a reasonable weight allocation strategy remains a challenge; dynamic weights may be affected by many factors such as the client's historical performance and trust, and it is difficult to weigh the uncertainty of different factors.
[0005] (3) The additional information transmission overhead is large: In order to achieve Byzantine robustness, some methods require additional information transmission, which increases the communication overhead. This contradicts one of the original intentions of federated learning, which is to reduce the amount of communication. Summary of the invention
[0006] In order to overcome the shortcomings of the prior art, the present invention aims to improve the versatility, adaptability and efficiency of Byzantine attack defense. By developing a more universal defense mechanism, it can effectively counter various attack modes, solving the limitation that traditional methods are only applicable to specific attack types. At the same time, by optimizing the dynamic weight allocation strategy, the present invention provides more accurate adaptability, comprehensively considers various factors, and enhances the robustness of the system. In addition, the present invention is committed to reducing unnecessary information transmission under the premise of ensuring Byzantine robustness to reduce communication overhead.
[0007] The solution adopted by the present invention is a federated learning Byzantine attack defense method based on gradient clustering and distance filtering, which includes the following steps:
[0008] Step 1: Calculate the instantaneous gradient g of each client in the client set B c and the cumulative gradient g a ;
[0009] Step 2: According to the instantaneous gradient and cumulative gradient of each client and other clients in the client set, the distance correlation and Wasserstein distance of each client are obtained, and the clients are filtered to obtain the client set D. k ;
[0010] Step 3: Calculate the recognition score of the client sub-gradient vector, dynamically assign weights to each client, and obtain the gradient for training the central server.
[0011] Furthermore, in step 1, the instantaneous gradient g of each client model in the client set B is calculated. c and the cumulative gradient g a The steps are:
[0012] B is the client set, where B = {B1, B2…B m ,…B N}, N is the number of client set B, B m For the mth client
[0013] Δ t represents the gradient vector obtained by the client set B passing to the server during the tth round of communication, Δ t ={Δ 1,t ,Δ 2,t ,…,Δ m,t ,...,Δ N,t}, the client's gradient vector dimension is l, Δ m,t represents the gradient vector of the mth client in the tth round of communication, where 1≤m≤N;
[0014] W i is the parameter of the central server’s model in round i;
[0015] Calculate the instantaneous gradient g of each client model in the client set B c , as shown in formula (1):
[0016]
[0017] h represents the sliding window size of the communication round number. When calculating the instantaneous gradient of the client, the gradient vector of the client from the th round to the tth round is taken. The average value of the gradient vector of the client h rounds is the instantaneous gradient g. c ; t represents the current communication round number;
[0018] The DBSCAN method based on density-based noise spatial clustering algorithm is used to classify the instantaneous gradient g of each client. c The client set B is clustered into two categories, and the client set B is classified according to the instantaneous gradient g cIn the two clustered categories, the number of elements in each category is determined, and the category with the largest number of elements is recorded as the client set clustered_c;
[0019] Calculate the cumulative gradient g for each client in the client set B a , as shown in formula (2):
[0020]
[0021] The DBSCAN method based on density-based noise spatial clustering algorithm is used to classify the accumulated gradient g of each client. a The client set B is clustered into two categories, and the client set B is classified according to the instantaneous gradient g a Of the two clustered categories, the one with the largest number of elements is recorded as the client set clustered_a;
[0022] Take the intersection of clustered_c and clustered_a to get clustered.
[0023] Furthermore, in step 2, the distance correlation and Wasserstein distance of each client are obtained according to the instantaneous gradient and cumulative gradient of each client and other clients in the client set, and the clients are filtered to obtain the client set D k The steps are:
[0024] Calculate the instantaneous gradient g between each client in the client set B and the N-1 clients in the set except itself c The distance correlation DC between them is summed up to get the instantaneous distance correlation of each client. The set of instantaneous distance correlations of N clients is dc c , dc c The size of the set is N;
[0025] Calculate the instantaneous gradient g between each client in the client set B and the N-1 clients in the set except itself c The Wasserstein distance WD between them is summed up to get the instantaneous Wasserstein distance of each client. The set of instantaneous Wasserstein distances of N clients is wd c , wd c The size of the set is N;
[0026] Calculate the cumulative gradient g of each client in the client set B and the N-1 clients in the set except itself a The distance correlation DC between them is summed up to get the cumulative gradient correlation of each client. The set of cumulative gradient correlations of N clients is dc a , dca The size of the set is N;
[0027] Calculate the cumulative gradient g of each client in the client set B and the N-1 clients in the set except itself a The Wasserstein distance WD between them is summed up to get the cumulative gradient Wasserstein distance of each client. The set of N client cumulative gradient Wasserstein distances is wd a , wd a The size of the set is N;
[0028] By using formula (3), dc c 、wd c 、dc a and wd a The corresponding elements in the set are calculated to obtain the distance d of each client in set B i , the distance d of all clients i The set is d, and the elements in set d are sorted from large to small to get set d′, from which the distance d with a ratio of k is selected. i The smallest client, get the client set D k ;
[0029] d i =wd ci +wd ai -dc ci -dc ai (3)
[0030] D k ={d j ∈d′:j∈[n]}, n=N×k (4)
[0031] where d j Indicates that it belongs to D k The jth client of D k The number of collection clients.
[0032] Furthermore, in step 3, the recognition score of the client sub-gradient vector is calculated, and the step of dynamically assigning a weight to each client is as follows:
[0033] Set client D k The instantaneous gradient g of each client in c Randomly split into p sub-gradients J c1 ,…,J cp , the client set D k The cumulative gradient g of each client in a Randomly split into p sub-gradients J a1 ,…,J ap, the dimension of each sub-gradient does not exceed l / p; the instantaneous gradient g of the jth client cj and the cumulative gradient g aj The sub-gradient of is expressed as formula (5):
[0034]
[0035] in represents the instantaneous gradient g of the jth client cj The qth sub-gradient vector and the jth client accumulated gradient g aj The qth sub-gradient vector of ;
[0036] By using formulas (6) and (7), the client set D k The instantaneous gradient g of each client in cj and the cumulative gradient g aj The sub-gradients are robustly aggregated to obtain the aggregated sub-gradient vector and
[0037]
[0038] Where M represents the median operation in the set; and Represent the qth robust gradient of the instantaneous gradient and the cumulative gradient respectively;
[0039] The client set D is calculated by formulas (8), (9), (10) and (11): k The recognition score of each client in:
[0040]
[0041] represents the recognition score of the qth sub-gradient of the instantaneous gradient of the jth client, represents the recognition score of the qth sub-gradient of the jth client cumulative gradient, represents the recognition score of the qth sub-gradient of the jth client, s j Represents the client set D k The recognition score of the jth client in ; WD() represents the Wasserstein distance between calculated gradients, and DC() represents the distance correlation between calculated gradients;
[0042] Assign a weight w to each client based on the recognition score j , as shown in formula (12):
[0043]
[0044] α represents the growth coefficient; wj represents the weight of the jth client; e is a natural constant;
[0045] Finally, the gradient g is calculated by formula (13): t :
[0046]
[0047] Gradient g t Used for training on the central server.
[0048] Furthermore, in step 2, the value of k is between 0 and 1.
[0049] Furthermore, in step three, the value of α is -1.
[0050] The beneficial effects of the present invention are as follows: by using different gradient information of the client, by identifying and screening the instantaneous gradient and the cumulative gradient, it is possible to have a defensive effect on different types of Byzantine attacks to adapt to different types of attacks. Distance correlation and Wasserstein distance are used to characterize the correlation between clients. Distance correlation and Wasserstein distance have wider applicability and better performance than cosine distance and Euclidean distance in terms of nonlinear relationships, scale invariance, outlier robustness, distribution comparison and high-dimensional data processing. When facing different types of Byzantine attacks, such as model poisoning and data poisoning, they all show strong robustness and high accuracy, especially in the case of medium and low proportions of attacks, their performance is significantly better than other mainstream defense strategies, showing higher recognition accuracy and stability. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 The overall framework diagram of the Byzantine attack defense of the present invention. DETAILED DESCRIPTION
[0052] The present invention is further described below in conjunction with the accompanying drawings and embodiments.
[0053] The federated learning Byzantine attack defense method based on gradient clustering and distance filtering includes the following steps:
[0054] Step 1: Calculate the instantaneous gradient g of each client model in the client set B c and the cumulative gradient g a ;
[0055] B is the client set, where B = {B1, B2…B m ,…B N}, N is the number of client set B, B m For the mth client
[0056] Δ trepresents the gradient vector obtained by the client set B passing to the server during the tth round of communication, Δ t ={Δ 1,t ,Δ 2,t ,…,Δ m,t ,...,Δ N,t}, the client's gradient vector dimension is l, Δ m,t represents the gradient vector of the mth client in the tth round of communication, where 1≤m≤N;
[0057] W i is the parameter of the central server’s model in round i;
[0058] Calculate the instantaneous gradient g of each client model in the client set B c , as shown in formula (1):
[0059]
[0060] h represents the sliding window size of the communication round number. When calculating the instantaneous gradient of the client, the gradient vector of the client from the th round to the tth round is taken. The average value of the gradient vector of the client h rounds is the instantaneous gradient g. c ; t represents the current communication round number;
[0061] The DBSCAN method based on density-based noise spatial clustering algorithm is used to classify the instantaneous gradient g of each client. c The client set B is clustered into two categories, and the client set B is classified according to the instantaneous gradient g c In the two clustered categories, the number of elements in each category is determined, and the category with the largest number of elements is recorded as the client set clustered_c;
[0062] Calculate the cumulative gradient g for each client in the client set B a , as shown in formula (2):
[0063]
[0064] The DBSCAN method based on density-based noise spatial clustering algorithm is used to classify the accumulated gradient g of each client. a The client set B is clustered into two categories, and the client set B is classified according to the instantaneous gradient g a Of the two clustered categories, the one with the largest number of elements is recorded as the client set clustered_a;
[0065] Take the intersection of clustered_c and clustered_a to get clustered;
[0066] Step 2: According to the instantaneous gradient and cumulative gradient of each client and other clients in the client set, the distance correlation and Wasserstein distance of each client are obtained, and the clients are filtered to obtain the client set D. k ;
[0067] Calculate the instantaneous gradient g between each client in the client set B and the N-1 clients in the set except itself c The distance correlation DC between them is summed up to get the instantaneous distance correlation of each client. The set of instantaneous distance correlations of N clients is dc c , dc c The size of the set is N;
[0068] Calculate the instantaneous gradient g between each client in the client set B and the N-1 clients in the set except itself c The Wasserstein distance WD between them is summed up to get the instantaneous Wasserstein distance of each client. The set of instantaneous Wasserstein distances of N clients is wd c , wd c The size of the set is N;
[0069] Calculate the cumulative gradient g of each client in the client set B and the N-1 clients in the set except itself a The distance correlation DC between them is summed up to get the cumulative gradient correlation of each client. The set of cumulative gradient correlations of N clients is dc a , dc a The size of the set is N;
[0070] Calculate the cumulative gradient g of each client in the client set B and the N-1 clients in the set except itself a The Wasserstein distance WD between them is summed up to get the cumulative gradient Wasserstein distance of each client. The set of N client cumulative gradient Wasserstein distances is wd a , wd a The size of the set is N;
[0071] By using formula (3), dc c 、wd c 、dc a and wd a The corresponding elements in the set are calculated to obtain the distance d of each client in set B i , the distance d of all clients i The set is d, and the elements in set d are sorted from large to small to get set d′, from which the distance d with a ratio of k is selected. iThe smallest client, get the client set D k ; k is a hyperparameter, with a value between 0 and 1;
[0072] d i =wd ci +wd ai -dc ci -dc ai (16)
[0073] D k ={d j ∈d′:j∈[n]}, n=N×k (17)
[0074] where d j Indicates that it belongs to D k The jth client of D k The number of collection clients;
[0075] Step 3: Calculate the recognition score of the client sub-gradient vector and dynamically assign weights to each client;
[0076] Set client D k The instantaneous gradient g of each client in c Randomly split into p sub-gradients J c1 ,…,J cp , the client set D k The cumulative gradient g of each client in a Randomly split into p sub-gradients J a1 ,…,J ap , the dimension of each sub-gradient does not exceed l / p; the instantaneous gradient g of the jth client cj and the cumulative gradient g aj The sub-gradient of is expressed as formula (5):
[0077]
[0078] in represents the instantaneous gradient g of the jth client cj The qth sub-gradient vector and the jth client accumulated gradient g aj The qth sub-gradient vector of ;
[0079] By using formulas (6) and (7), the client set D k The instantaneous gradient g of each client in cj and the cumulative gradient g aj The sub-gradients are robustly aggregated to obtain the aggregated sub-gradient vector and
[0080]
[0081] Where M represents the median operation in the set; and Represent the qth robust gradient of the instantaneous gradient and the cumulative gradient respectively;
[0082] The client set D is calculated by formulas (8), (9), (10) and (11): k The recognition score of each client in:
[0083]
[0084] represents the recognition score of the qth sub-gradient of the instantaneous gradient of the jth client, represents the recognition score of the qth sub-gradient of the jth client cumulative gradient, represents the recognition score of the qth sub-gradient of the jth client, s j Represents the client set D k The recognition score of the jth client in ; WD() represents the Wasserstein distance between calculated gradients, and DC() represents the distance correlation between calculated gradients;
[0085] Assign a weight w to each client based on the recognition score j , as shown in formula (12):
[0086]
[0087] α represents the growth coefficient, which is taken as -1 here; w j represents the weight of the jth client; e is a natural constant;
[0088] Finally, the gradient g is calculated by formula (13): t :
[0089]
[0090] Gradient g t Used for training on the central server.
[0091] In the present invention, step one takes into account different gradient information of the client, and by identifying and screening the instantaneous gradient and the cumulative gradient, it can have a defensive effect on different types of Byzantine attacks to adapt to different types of attacks.
[0092] Step 2 uses distance correlation and Wasserstein distance to characterize the correlation between clients. Distance correlation and Wasserstein distance have wider applicability and better performance than cosine distance and Euclidean distance in terms of nonlinear relationships, scale invariance, outlier robustness, distribution comparison, and high-dimensional data processing. These characteristics make them more suitable for measuring the similarity between clients in areas such as federated learning.
[0093] Step 3 dynamically assigns weights to each client by calculating the recognition score of the client sub-gradient vector. The present invention divides the gradient into different sub-gradients, fully considers the heterogeneity between gradients, can detect abnormal gradients, and can also effectively aggregate the gradients of non-independent, isotropic and heterogeneous data sets.
[0094] The following is the robustness and accuracy of the present invention when facing different types of Byzantine attacks (model poisoning and data poisoning) when applied to different data sets:
[0095] A. Federated Extended Handwritten Numbers and Letters Dataset:
[0096] In the absence of attacks, the accuracy of the GCDF (Gradient Clustering and Distance Filtering) method of this patent is 83.25%, which is comparable to other methods. Under model poisoning attacks, GCDF performs best in all attack ratios (10%, 20%, 30%), especially at 10% and 20% attack ratios, with accuracy rates of 83.35% and 83.02% respectively. Under data poisoning attacks, the GCDF method also maintains a high accuracy rate, especially at 10% and 30% attack ratios, reaching 82.93% and 82.15% respectively.
[0097] B. Canadian Institute for Advanced Research 10-category image dataset:
[0098] In the absence of attacks, the accuracy of the GCDF method is 65.65%, the highest among all methods. Under model poisoning attacks, GCDF has the highest accuracy at 10% and 20% attack ratios, which are 65.35% and 64.92% respectively. Under data poisoning attacks, the GCDF method performs well at 10%, 30%, and 50% attack ratios, with accuracy rates of 61.48%, 59.81%, and 59.33% respectively.
[0099] C. Canadian Institute for Advanced Research 100-category image dataset:
[0100] In the absence of attacks, the GCDF method has an accuracy of 47.76%, which is a good performance. Under model poisoning attacks, GCDF has the highest accuracy at 10% and 20% attack ratios, which are 47.14% and 45.32%, respectively. Under data poisoning attacks, the GCDF method performs well at 30% and 50% attack ratios, with accuracy rates of 44.93% and 44.73%, respectively.
[0101] In summary, the GCDF method shows strong robustness and high accuracy when facing different types of Byzantine attacks (model poisoning and data poisoning). Especially in the case of medium and low proportion attacks, its performance is significantly better than other mainstream defense strategies, showing higher recognition accuracy and stability.
[0102] Table 1
[0103]
[0104] The table shows that under three data sets, the accuracy of the two major types of attacks, model poisoning and data poisoning, with different Byzantine node ratios, is basically better than the existing methods.
Claims
1. A federated learning Byzantine attack defense method based on gradient clustering and distance filtering, characterized in that: The steps include: Step 1: Calculate the instantaneous gradient g of each client in the client set B c and the cumulative gradient g a ; Step 2: According to the instantaneous gradient and cumulative gradient of each client and other clients in the client set, the distance correlation and Wasserstein distance of each client are obtained, and the clients are filtered to obtain the client set D. k ; Step 3: Calculate the recognition score of the client sub-gradient vector, dynamically assign weights to each client, and obtain the gradient for training the central server.
2. According to claim 1, a federated learning Byzantine attack defense method based on gradient clustering and distance filtering is characterized in that: In step 1, the instantaneous gradient g of each client model in the client set B is calculated c and the cumulative gradient g a The steps are: B is the client set, where B = {B1, B2…B m ,…B N }, N is the number of client set B, B m For the mth client Δ t represents the gradient vector obtained by the client set B passing to the server during the tth round of communication, Δ t ={Δ 1,t ,Δ 2,t ,…,Δ m,t ,...,Δ N,t }, the client's gradient vector dimension is l, Δ m,t represents the gradient vector of the mth client in the tth round of communication, where 1≤m≤N; W i is the parameter of the central server’s model in round i; Calculate the instantaneous gradient g of each client model in the client set B c , as shown in formula (1): h represents the sliding window size of the communication round number. When calculating the instantaneous gradient of the client, the gradient vector of the client from the th round to the tth round is taken. The average value of the gradient vector of the client h rounds is the instantaneous gradient g. c ; t represents the current communication round number; The DBSCAN method based on density-based noise spatial clustering algorithm is used to classify the instantaneous gradient g of each client. c The client set B is clustered into two categories, and the client set B is classified according to the instantaneous gradient g c In the two clustered categories, the number of elements in each category is determined, and the category with the largest number of elements is recorded as the client set clustered_c; Calculate the cumulative gradient g for each client in the client set B a , as shown in formula (2): The DBSCAN method based on density-based noise spatial clustering algorithm is used to classify the accumulated gradient g of each client. a The client set B is clustered into two categories, and the client set B is classified according to the instantaneous gradient g a Of the two clustered categories, the one with the largest number of elements is recorded as the client set clustered_a; Take the intersection of clustered_c and clustered_a to get clustered.
3. According to claim 1, a federated learning Byzantine attack defense method based on gradient clustering and distance filtering is characterized in that: In step 2, the distance correlation and Wasserstein distance of each client are obtained according to the instantaneous gradient and cumulative gradient of each client and other clients in the client set, and the clients are filtered to obtain the client set D k The steps are: Calculate the instantaneous gradient g between each client in the client set B and the N-1 clients in the set except itself c The distance correlation DC between them is summed up to get the instantaneous distance correlation of each client. The set of instantaneous distance correlations of N clients is dc c , dc c The size of the set is N; Calculate the instantaneous gradient g between each client in the client set B and the N-1 clients in the set except itself c The Wasserstein distance WD between them is summed up to get the instantaneous Wasserstein distance of each client. The set of instantaneous Wasserstein distances of N clients is wd c , wd c The size of the set is N; Calculate the cumulative gradient g of each client in the client set B and the N-1 clients in the set except itself a The distance correlation DC between them is summed up to get the cumulative gradient correlation of each client. The set of cumulative gradient correlations of N clients is dc a , dc a The size of the set is N; Calculate the cumulative gradient g of each client in the client set B and the N-1 clients in the set except itself a The Wasserstein distance WD between them is summed up to get the cumulative gradient Wasserstein distance of each client. The set of N client cumulative gradient Wasserstein distances is wd a , wd a The size of the set is N; By using formula (3), dc c 、wd c 、dc a and wd a The corresponding elements in the set are calculated to obtain the distance d of each client in set B i , the distance d of all clients i The set is d, and the elements in set d are sorted from large to small to get set d′, from which the distance d with a ratio of k is selected. i The smallest client, get the client set D k ; d i =wd ci +wd ai -dc ci -dc ai (3) D k ={d j ∈d′:j∈[n]}, n=N×k (4) where d j Indicates that it belongs to D k The jth client of D k The number of collection clients.
4. According to claim 1, a federated learning Byzantine attack defense method based on gradient clustering and distance filtering is characterized in that: Furthermore, in step 3, the recognition score of the client sub-gradient vector is calculated, and the step of dynamically assigning a weight to each client is as follows: Set client D k The instantaneous gradient g of each client in c Randomly split into p sub-gradients J c1 ,…,J cp , the client set D k The cumulative gradient g of each client in a Randomly split into p sub-gradients J a1 ,…,J ap , the dimension of each sub-gradient does not exceed l / p; the instantaneous gradient g of the jth client cj and the cumulative gradient g aj The sub-gradient of is expressed as formula (5): in represents the instantaneous gradient g of the jth client cj The qth sub-gradient vector and the jth client accumulated gradient g aj The qth sub-gradient vector of ; By using formulas (6) and (7), the client set D k The instantaneous gradient g of each client in cj and the cumulative gradient g aj The sub-gradients are robustly aggregated to obtain the aggregated sub-gradient vector and Where M represents the median operation in the set; and Represent the qth robust gradient of the instantaneous gradient and the cumulative gradient respectively; The client set D is calculated by formulas (8), (9), (10) and (11): k The recognition score of each client in: represents the recognition score of the qth sub-gradient of the instantaneous gradient of the jth client, represents the recognition score of the qth sub-gradient of the jth client cumulative gradient, represents the recognition score of the qth sub-gradient of the jth client, s j Represents the client set D k The recognition score of the jth client in ; WD() represents the Wasserstein distance between calculated gradients, and DC() represents the distance correlation between calculated gradients; Assign a weight w to each client based on the recognition score j , as shown in formula (12): α represents the growth coefficient; w j represents the weight of the jth client; e is a natural constant; Finally, the gradient g is calculated by formula (13): t : Gradient g t Used for training of central servers.
5. The method for defending against Byzantine attacks based on federated learning using gradient clustering and distance filtering according to claim 1, characterized in that: In step 2, the value of k is between 0 and 1.
6. The method for defending against Byzantine attacks based on federated learning based on gradient clustering and distance filtering according to claim 1, characterized in that: In the step 3, the value of α is -1.
Citation Information
Cited By
Federal learning-based Byzantine attack defense method, apparatus and device, and medium
CN120639433A
Federal learning based byzantine attack defense method, device, equipment and medium
CN120639433B
