Federal forgetting method and system based on projection gradient rise
By employing a federated forgetting method based on projected gradient ascent, the problems of privacy leakage and catastrophic forgetting in federated learning are addressed, achieving efficient data deletion and model consistency while ensuring data privacy and performance.
Patent Information
- Application Number
- CN202511424147.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-30
- Publication Date
- 2025-11-21
AI Technical Summary
In federated learning environments, existing machine forgetting techniques are ineffective at deleting personal data, leading to privacy breaches and security risks. Furthermore, gradient ascent methods are prone to catastrophic forgetting and lack active participation from the target client.
A federated forgetting method based on projection gradient ascent is adopted. A small-scale distillation dataset is generated through data distillation, a reference model is obtained and optimized by gradient ascent, and the forgetting client and the remaining client are trained by combining gradient descent algorithm. The central server performs model aggregation.
Reduce computational and communication overhead, ensure the forgetting model closely approximates the data distribution of other clients, prevent performance impact, and enable a flexible forgetting process.
Smart Images

Figure CN120996143A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of machine learning technology, and in particular to a federated forgetting method and system based on projective gradient ascent. Background Technology
[0002] As demand increases, machine learning relies on massive amounts of data to improve performance, but data privacy issues have also attracted widespread attention. Various countries have enacted data privacy regulations, such as the EU's General Data Protection Regulation (GDPR) and California's Consumer Privacy Act, granting users the "Right to be forgotten" (RTBF), allowing the deletion of their personal data. The implementation of the "Right to be forgotten" relies on "machine forgetting" technology to ensure the effective removal of personal data from training datasets and models, thereby enhancing data privacy and ethical practices.
[0003] To improve performance and efficiency while protecting user privacy, federated learning emerged. However, while the model-sharing mechanism of federated learning promotes data collaboration, it also introduces privacy leaks and security risks, such as obtaining data information through member inference and model flipping. Furthermore, attacks like data poisoning can affect model performance, requiring more stringent data protection and defense measures. Therefore, the technique of incorporating machine forgetting based on federated learning is called Federated Unlearning (FU).
[0004] However, simply deleting client data or using existing MU algorithms cannot achieve federated forgetting (FU). This is because in federated forgetting (FL), model training is completed through multiple iterations, and the information in parameter updates is forward-coupled. Traditional machine forgetting techniques designed for centralized machine learning are no longer applicable in the FL environment. Therefore, compared to machine forgetting schemes in non-distributed environments, knowledge permeation makes achieving performance consistency between the forgotten model and the retrained model in federated forgetting more difficult.
[0005] Existing forgetting (FU) methods are either implemented in a distributed manner or use gradient ascent. Distributed methods typically sacrifice storage space for effective forgetting, while gradient ascent takes a reverse approach. Instead of traditional gradient descent, it optimizes by rolling back the model and performing several rounds of fine-tuning to achieve FU. However, this reverse gradient descent approach is prone to catastrophic forgetting and does not involve the target client actively in the forgetting process. Summary of the Invention
[0006] The purpose of this invention is to provide a federated forgetting method and system based on projection gradient ascent, so as to solve the problems existing in the background technology.
[0007] To achieve the above objectives, this invention provides a federated forgetting method based on projective gradient ascent, comprising the following steps:
[0008] S1. Obtain information on N clients and the central server in the federation, and use the federation weighted aggregation learning algorithm to train the federation for T rounds;
[0009] S2. Obtain information on N clients and the central server in the federation, and perform data distillation on each client to obtain a small-scale distilled dataset;
[0010] S3. Obtain the federated global model ω obtained in S1. T And the local model of the forgotten client in round T-1 Obtaining the reference model ω on the forgotten client ref ;
[0011] S4. Obtain the reference model ω obtained in S3. ref Client i is located at ω ref The model parameters in an l2 norm sphere with a radius of δ are optimized.
[0012] S5. Obtain the federated global model ω obtained in S4. T-1 , will ω T-1 Send the training to all clients, and apply the opposite training strategy to the forgotten clients and the remaining clients;
[0013] S6, Obtain the forgotten client local model ω from S5. i and the remaining client-side local model ω j The central server re-aggregates the models:
[0014]
[0015] Wherein, λ is the influence factor of the forgetting model during aggregation, and λ is preferably 0.1;
[0016] S7. For the information on the N clients and the central server in the federation obtained in S1, repeat S5 and S6 until E rounds of training are completed, obtaining the final federated forgetting model ω. u .
[0017] Preferably, step S1 is as follows:
[0018] S11. Obtain the parameter ω of the current global model M on the central server. t The global model is sent to each client.
[0019] S12. Obtain the N clients in the federation and their corresponding datasets D. i ={(xi ,y i ), i∈[n i ]} and federated learning model M i Several rounds of deep learning training are performed on each client. Specifically, the training objective for each client is:
[0020]
[0021] Where, θ i Represents the model parameters on client i, (x i ,y i )~D i Let L be the dataset sample for client i. L(·) is the loss function, and M is the loss function. i (x i ,θ i ) is the model's input x i The predicted output at that time.
[0022] S13. Obtain the updated local model for each client obtained in step S12. Each client sends the updated local model to the central server.
[0023] S14. Obtain the updated local model obtained in step S13. We aggregate them using a weighted average to obtain a new global model ω. t+1 ω t+1 The calculation is defined as follows:
[0024]
[0025] in, That is, p i It represents the proportion of data from the i-th client in the entire FL system.
[0026] S15. For the information on the N clients and the central server in the federation obtained in step S1, repeat steps S11 to S14 until T rounds of training are completed, and the final federated learning model ω is obtained. T .
[0027] In this step, the federated learning model is FLNet, and the loss function is set to cross-entropy loss. The hyperparameters of the FLNet model were set as follows: learning rate of 0.01, batch size of 16, and stochastic gradient descent optimizer.
[0028] Preferably, step S2 is as follows:
[0029] S21. Obtain the N clients in the federation and their corresponding datasets D. i={(x i ,y i ), i∈[n i Random initialization and D i Generate samples of the same size
[0030] S22. Obtain the datasets D corresponding to the N clients in the federation. i ={(x i ,y i ), i∈[n i And step S12 yields the federated learning model M. i and the product generated in step S21 Copy the model to obtain M' i Calculate M respectively i and gradient information and
[0031] S23. Obtain the gradient information obtained in step S21. and Update using gradient descent. Specifically:
[0032]
[0033] in, These represent the number of steps and the learning rate, respectively, and d(.;.) is... and The distance between them is a function, and here we use Euclidean distance as the measure.
[0034] S24, Obtain the result from S22 Obtain the copy model M' from step S21 i ,use For M' i The update will be performed as follows:
[0035]
[0036] Where, η θ This is the updated learning rate. Step S2 is performed on all clients for T rounds of gradient matching, gradually increasing the learning rate. The quality.
[0037] Preferably, the reference model ω in S3 ref as follows:
[0038]
[0039] Preferably, step S4 is as follows:
[0040] S41. Obtain the reference model ω obtained in step S3. ref An l2 norm sphere with an surrounding radius of δ. Specifically, the l2 norm sphere can be represented by the following formula:
[0041]
[0042] S42. Obtain the l2 norm sphere Ω obtained in step S41, and obtain the given step size η. u And the data D obtained from client i in step S1 i The model is updated using the projected gradient ascent method. Specifically, the projected gradient ascent method can be expressed as:
[0043]
[0044] in, Let be the projection operator on Ω. It is the loss function relative to D i The gradient of the projected gradient. In this step, the ascent of the projected gradient ends when the distance between the forgotten model and the current client model is greater than a predetermined threshold τ. Preferably, τ is set to 2.2.
[0045] S43. Obtain the forgetting model ω obtained in step S42 and send ω back to the central server.
[0046] Preferably, the training algorithm for forgotten clients and remaining clients in S5 is as follows:
[0047] For the forgotten client, retrieve data D from step S1. i The gradient ascent algorithm is used for several rounds of training.
[0048]
[0049] For the remaining clients, obtain the data D for each remaining client in step S1. j Perform several rounds of training using the gradient descent algorithm:
[0050]
[0051] Where η is the learning rate, preferably 0.01.
[0052] A federated forgetting system based on projective gradient ascent includes:
[0053] The first module obtains information about N clients and the central server in the federation, and uses the federation weighted aggregation learning algorithm to train the federation for T rounds.
[0054] The second module obtains information about N clients and the central server in the federation, performs data distillation on each client, and obtains a small-scale distilled dataset.
[0055] The third module obtains the federated global model ω obtained from the first module. T And the local model of the forgotten client in round T-1 Obtaining the reference model ω on the forgotten client ref Specifically:
[0056]
[0057] The fourth module obtains the reference model ω obtained in the third module. ref Client i is located at ω ref The model parameters in an l2 norm sphere with a radius of δ are optimized.
[0058] The fifth module obtains the federated global model ω from the fifth module. T-1 , will ω T-1 Send the training to all clients, and apply the opposite training strategy to the forgotten clients and the remaining clients.
[0059] Specifically, for the forgotten client, data is retrieved from the second module. Perform several rounds of training using the gradient ascent algorithm:
[0060]
[0061] For the remaining clients, retrieve the data for each remaining client in the second module. Perform several rounds of training using the gradient descent algorithm:
[0062]
[0063] Where η is the learning rate, preferably 0.01.
[0064] The sixth module obtains the forgotten client local model ω from the fifth module. i and the remaining client-side local model ω j The central server re-aggregates the models:
[0065]
[0066] Wherein, λ is the influence factor of the forgetting model during aggregation, and λ is preferably 0.1.
[0067] The seventh module, based on the information about the N clients and the central server in the federation obtained in the first module, repeats the fifth and sixth modules until E rounds of training are completed, resulting in the final federated forgetting model ω. u.
[0068] Therefore, the present invention employs the above-mentioned federated forgetting method and system based on projection gradient ascent, which has the following beneficial effects:
[0069] (1) Since the gradient information obtained by the local model in step S2 is reused, a small-scale distillation dataset is generated for each client, and the computational overhead is reduced in the subsequent fine-tuning process by using the distillation data.
[0070] (2) Step S3 was used to solve the reference model. By using the average value of other clients as the reference model, it can be ensured that the forgotten model is close enough to the data distribution of other clients that have been effectively learned. At the same time, by solving the reference model on the forgotten client instead of on the central server, additional communication overhead is avoided.
[0071] (3) Since steps S4 to S6 are used to fine-tune the federated forgetting model, the forgetting client is included in the fine-tuning, the target client model is adopted by the gradient ascent algorithm, and the λ weight is introduced to multiply the target client model to reduce its contribution and prevent it from affecting the performance of the remaining dataset.
[0072] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0073] Figure 1 This is a flowchart of the federal forgetting process in an embodiment of the present invention;
[0074] Figure 2 This is an overall framework diagram in an embodiment of the present invention;
[0075] Figure 3 This is a schematic diagram of federated learning training in an embodiment of the present invention;
[0076] Figure 4 This is a diagram illustrating the overall framework for obtaining the reference model and projected gradient ascent federation in an embodiment of the present invention.
[0077] Figure 5 This is a diagram of the overall framework for training the federated system in an embodiment of the present invention. Detailed Implementation
[0078] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0079] The basic idea of this embodiment is as follows: after training the federation of N clients for T rounds, client i∈[N] requests to leave the federation and hopes to eliminate the influence of its entire local data from the federation. After obtaining the reference model by averaging the remaining clients, the projection gradient ascent is applied to obtain the local model of each client. The central server aggregates the models of multiple clients and distributes them to each client. Finally, the model is fine-tuned by combining it with the model of the target client to obtain the final forgetting model.
[0080] Please see Figures 1-5 A federated forgetting method based on projected gradient ascent includes the following steps:
[0081] S1. Obtain information on N clients and the central server in the federation, and use the federation weighted aggregation learning algorithm to train the federation for T rounds.
[0082] S11. Obtain the parameter ω of the current global model M on the central server. t The global model is sent to each client.
[0083] S12. Obtain the N clients in the federation and their corresponding datasets D. i ={(x i ,y i ), i∈[n i ]} and Federated Learning Model M i Several rounds of deep learning training are performed on each client. Specifically, the training objective for each client is:
[0084]
[0085] Where, θ i Represents the model parameters on client i, (x i ,y i )~D i Let L be the dataset sample for client i. L(·) is the loss function, and M is the loss function. i (x i ,θ i ) is the model's input x i The predicted output at that time.
[0086] S13. Obtain the updated local model for each client obtained in step S12. Each client sends the updated local model to the central server.
[0087] S14. Obtain the updated local model obtained in step S13. We aggregate them using a weighted average to obtain a new global model ω. t+1 ω t+1The calculation is defined as follows:
[0088]
[0089] in, That is, p i It represents the proportion of data from the i-th client in the entire FL system.
[0090] S15. For the information on the N clients and the central server in the federation obtained in step S1, repeat steps S11 to S14 until T rounds of training are completed, and the final federated learning model ω is obtained. T .
[0091] In this step, the federated learning model is FLNet, and the loss function is set to cross-entropy loss. The hyperparameters of the FLNet model were set as follows: learning rate of 0.01, batch size of 16, and stochastic gradient descent optimizer.
[0092] S2. Obtain information on N clients and the central server in the federation, perform data distillation on each client, and obtain a small-scale distilled dataset.
[0093] S21. Obtain the N clients in the federation and their corresponding datasets D. i ={(x i ,y i ), i∈[n i Random initialization and D i Generate samples of the same size
[0094] S22. Obtain the datasets D corresponding to the N clients in the federation. i ={(x i ,y i ), i∈[n i And step S12 yields the federated learning model M. i and the product generated in step S21 Copy the model to obtain M' i Calculate M respectively i and gradient information and
[0095] S23. Obtain the gradient information obtained in step S21. and Update using gradient descent. Specifically:
[0096]
[0097] in, These represent the number of steps and the learning rate, respectively, and d(.;.) is... and The distance between them is a function, and here we use Euclidean distance as the measure.
[0098] S24, Obtain the result from S22 Obtain the copy model M' from step S21 i ,use For M' i The update will be performed as follows:
[0099]
[0100] Where, η θ This is the updated learning rate. Step S2 is performed on all clients for T rounds of gradient matching, gradually increasing the learning rate. The quality.
[0101] S3. Obtain the federated global model ω obtained in S1. T And the local model of the forgotten client in round T-1 Obtaining the reference model ω on the forgotten client ref :
[0102]
[0103] S4. Obtain the reference model ω obtained in S3. ref Client i is located at ω ref The model parameters in an l2 norm sphere with a radius of δ are optimized.
[0104] S41. Obtain the reference model ω obtained in step S3. ref An l2 norm sphere with an surrounding radius of δ. Specifically, the l2 norm sphere can be represented by the following formula:
[0105]
[0106] S42. Obtain the l2 norm sphere Ω obtained in step S41, and obtain the given step size η. u And the data D obtained from client i in step S1 i The model is updated using the projected gradient ascent method. Specifically, the projected gradient ascent method can be expressed as:
[0107]
[0108] in, Let be the projection operator on Ω. It is the loss function relative to D iThe gradient of the projected gradient. In this step, the ascent of the projected gradient ends when the distance between the forgotten model and the current client model is greater than a predetermined threshold τ. Preferably, τ is set to 2.2.
[0109] S43. Obtain the forgetting model ω obtained in step S42 and send ω back to the central server.
[0110] S5. Obtain the federated global model ω obtained in S4. T-1 , will ω T-1 Send the training to all clients, and apply the opposite training strategy to the forgotten clients and the remaining clients.
[0111] For the forgotten client, retrieve data D from step S1. i The gradient ascent algorithm is used for several rounds of training.
[0112]
[0113] For the remaining clients, obtain the data D for each remaining client in step S1. j Perform several rounds of training using the gradient descent algorithm:
[0114]
[0115] Where η is the learning rate, preferably 0.01.
[0116] S6, Obtain the forgotten client local model ω from S5. i and the remaining client-side local model ω j The central server re-aggregates the models:
[0117]
[0118] Wherein, λ is the influence factor of the forgetting model during aggregation, and λ is preferably 0.1.
[0119] S7. For the information on the N clients and the central server in the federation obtained in S1, repeat S5 and S6 until E rounds of training are completed, obtaining the final federated forgetting model ω. u .
[0120] A federated forgetting system based on projective gradient ascent includes:
[0121] The first module obtains information about N clients and the central server in the federation, and uses the federation weighted aggregation learning algorithm to train the federation for T rounds.
[0122] The second module obtains information about N clients and the central server in the federation, performs data distillation on each client, and obtains a small-scale distilled dataset.
[0123] The third module obtains the federated global model ω obtained from the first module. T And the local model of the forgotten client in round T-1 Obtaining the reference model ω on the forgotten client ref Specifically:
[0124]
[0125] The fourth module obtains the reference model ω obtained in the third module. ref Client i is located at ω ref The model parameters in an l2 norm sphere with a radius of δ are optimized.
[0126] The fifth module obtains the federated global model ω from the fifth module. T-1 , will ω T-1 Send the training to all clients, and apply the opposite training strategy to the forgotten clients and the remaining clients.
[0127] Specifically, for the forgotten client, data is retrieved from the second module. Perform several rounds of training using the gradient ascent algorithm:
[0128]
[0129] For the remaining clients, retrieve the data for each remaining client in the second module. Perform several rounds of training using the gradient descent algorithm:
[0130]
[0131] Where η is the learning rate, preferably 0.01.
[0132] The sixth module obtains the forgotten client local model ω from the fifth module. i and the remaining client-side local model ω j The central server re-aggregates the models:
[0133]
[0134] Wherein, λ is the influence factor of the forgetting model during aggregation, and λ is preferably 0.1.
[0135] The seventh module, based on the information about the N clients and the central server in the federation obtained in the first module, repeats the fifth and sixth modules until E rounds of training are completed, resulting in the final federated forgetting model ω. u .
[0136] Therefore, this invention employs the aforementioned federated forgetting method and system based on projective gradient ascent to address the technical problem that knowledge permeation in distributed environments makes it difficult to apply machine forgetting methods for non-distributed environments to federated learning. Simultaneously, it addresses the technical challenge of catastrophic forgetting during gradient ascent in federated forgetting, providing a more flexible solution.
[0137] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A federated forgetting method based on projective gradient ascent, characterized in that, Includes the following steps: S1. Obtain information on N clients and the central server in the federation, and use the federation weighted aggregation learning algorithm to train the federation for T rounds; S2. Obtain information on N clients and the central server in the federation, and perform data distillation on each client to obtain a small-scale distilled dataset; S3. Obtain the federated global model ω obtained in S1. T And the local model of the forgotten client in round T-1 Obtaining the reference model ω on the forgotten client ref ; S4. Obtain the reference model ω obtained in S3. ref Client i is located at ω ref The model parameters in an l2 norm sphere with a radius of δ are optimized. S5. Obtain the federated global model ω obtained in S4. T-1 , will ω T-1 Send the training to all clients, and apply the opposite training strategy to the forgotten clients and the remaining clients; S6. Obtain the forgotten client local model ω obtained in S5. i and the remaining client-side local model ω j The central server re-aggregates the models: Wherein, λ is the influence factor of the forgetting model during aggregation, and λ is preferably 0.1; S7. For the information on the N clients and the central server in the federation obtained in S1, repeat S5 and S6 until E rounds of training are completed, obtaining the final federated forgetting model ω. u .
2. The federated forgetting method based on projective gradient ascent according to claim 1, characterized in that, Step S1 is as follows: S11. Obtain the parameter ω of the current global model M on the central server. t Send the global model to each client; S12. Obtain the N clients in the federation and their corresponding datasets D. i ={(x i ,y i ), i∈[n i ]} and federated learning model M i Perform several rounds of deep learning training on each client; the training objective for each client is: Where, θ i Represents the model parameters on client i, (x i ,y i )~D i Let L be the dataset sample of client i; L(·) is the loss function, and M is the dataset sample of client i. i (x i ,θ i ) is the model's input x i The predicted output at that time; S13. Obtain the updated local model for each client obtained in step S12. Each client sends the updated local model to the central server; S14. Obtain the updated local model obtained in step S13. We aggregate them using a weighted average to obtain a new global model ω. t+1 ;ω t+1 The calculation is as follows: in, p i It represents the proportion of data from the i-th client in the entire FL system; S15. For the information on the N clients and the central server in the federation obtained in step S1, repeat steps S11 to S14 until T rounds of training are completed, and the final federated learning model ω is obtained. T .
3. The federated forgetting method based on projective gradient ascent according to claim 2, characterized in that, The steps in S2 are as follows: S21. Obtain the N clients in the federation and their corresponding datasets D. i ={(x i ,y i ), i∈[n i Random initialization and D i Generate samples of the same size S22. Obtain the datasets D corresponding to the N clients in the federation. i ={(x i ,y i ), i∈[n i ]} and the federated learning model M obtained in step S12 i and the product generated in step S21 M' is obtained by copying the federated learning model. i Calculate M respectively i and gradient information and S23. Obtain the gradient information obtained in step S21. and Update using gradient descent. in, These represent the number of steps and the learning rate, respectively, and d(.;.) is... and A function of the distance between them; S24, Obtain the result from S22 Obtain the copy model M' from step S21 i ,use For M' i The update will be performed as follows: Where, η θ This involves updating the learning rate; performing step S2 for T rounds of gradient matching across all clients, gradually increasing the learning rate. The quality.
4. The federated forgetting method based on projective gradient ascent according to claim 1, characterized in that, Reference model ω in S3 ref as follows:
5. The federated forgetting method based on projective gradient ascent according to claim 4, characterized in that, The S4 steps are as follows: S41. Obtain the reference model ω obtained in step S3. ref An l2 norm sphere with a radius of δ is defined by the following formula: S42. Obtain the l2 norm sphere Ω obtained in step S41, and obtain the given step size η. u And the data D obtained from client i in step S1 i The model is updated using the projected gradient ascent method, which is expressed as: in, For projection operators on Ω; It is the loss function relative to D i The gradient of the projected gradient ends when the distance between the forgotten model and the current client model is greater than a predetermined threshold τ, where τ is 2.
2. S43. Obtain the forgetting model ω obtained in step S42 and send ω back to the central server.
6. The federated forgetting method based on projective gradient ascent according to claim 5, characterized in that: The training algorithm for forgotten clients and remaining clients in S5 is as follows: For the forgotten client, retrieve data D from step S1. i The gradient ascent algorithm is used for several rounds of training. For the remaining clients, obtain the data D for each remaining client in step S1. j Perform several rounds of training using the gradient descent algorithm: Where η is the learning rate, with a value of 0.
01.
7. A federated forgetting system based on projective gradient ascent, employing the federated forgetting method based on projective gradient ascent as described in any one of claims 1-6, characterized in that, include: The first module obtains information about N clients and the central server in the federation, and uses the federation weighted aggregation learning algorithm to train the federation for T rounds. The second module obtains information about N clients and the central server in the federation, performs data distillation on each client, and obtains a small-scale distilled dataset. The third module obtains the federated global model ω obtained from the first module. T And the local model of the forgotten client in round T-1 Obtaining the reference model ω on the forgotten client ref : The fourth module obtains the reference model ω obtained in the third module. ref Client i is located at ω ref The model parameters in an l2 norm sphere with a radius of δ are optimized. The fifth module obtains the federated global model ω from the fifth module. T-1 , will ω T-1 Send the training to all clients, and apply the opposite training strategy to the forgotten clients and the remaining clients; For the forgotten client, retrieve the data from the second module. Perform several rounds of training using the gradient ascent algorithm: For the remaining clients, retrieve the data for each remaining client in the second module. Perform several rounds of training using the gradient descent algorithm: Where η is the learning rate, with a value of 0.01; The sixth module obtains the forgotten client local model ω from the fifth module. i and the remaining client-side local model ω j The central server re-aggregates the models: Wherein, λ is the influence factor of the forgetting model during aggregation, and λ is 0.1; The seventh module, based on the information about the N clients and the central server in the federation obtained in the first module, repeats the fifth and sixth modules until E rounds of training are completed, resulting in the final federated forgetting model ω. u .
Citation Information
Cited By
Edge end participating in cloud-edge federated learning, control method and cloud-edge federated learning system
CN121644601A
Edge terminal participating in cloud-edge federated learning and operation method, and cloud-edge federated learning system
CN121644601B
