Asynchronous federated learning method and system based on dynamic contribution awareness and staleness adaptation

Abstract

The application discloses an asynchronous federated learning method and system based on dynamic contribution perception and obsolescence adaptation, and the method comprises the following steps: after a server broadcasts a latest global model and a contribution degree vector, corresponding contribution degrees are received; it is judged whether the received contribution degrees meet training conditions, if yes, the latest global model is downloaded for training, and if not, an old version global model is continuously used for training; a local optimization objective function is constructed by fusing a loss function, a model deviation penalty and an obsolescence inhibition term; a dynamic learning rate is calculated; a gradient descent is performed to update a local model, and the local model is sent to the server, so that the server performs weighted aggregation on the local model by using a dynamic aggregation weight after receiving more than or equal to a preset number of client local models, updates client contribution degrees, and generates a contribution degree vector by normalization. The application takes into account the communication efficiency and the model accuracy, is suitable for large-scale distributed machine learning scenes, and has the advantages of strong robustness and good adaptability.

Description

TECHNICAL FIELD

[0001] The application relates to an asynchronous federated learning method and system based on dynamic contribution awareness and staleness adaptation, and belongs to the technical field of federated learning. BACKGROUND

[0002] Due to the characteristics of distributed computing and privacy protection, federated learning has been widely applied in the fields of intelligent devices, Internet of Things and medical health. Specifically, federated learning allows multiple clients to train models locally and aggregate model updates from each client through a server, thereby optimizing the global model. However, the traditional synchronous federated learning method has some defects. In the traditional synchronous federated learning, the center server must wait for all clients to complete local training and upload model updates, so that some offline or less computationally capable clients will slow down the aggregation speed of the global model.

[0003] On this basis, the asynchronous federated learning method emerged as the times require. Asynchronous federated learning allows the server to not wait for all clients to complete training, but can aggregate immediately after receiving model updates from some clients. This method can significantly improve training efficiency, especially in the case of a large number of clients or large network latency. However, existing asynchronous federated learning methods still have some problems. First, in the asynchronous update system, local models from different clients are trained, updated and uploaded asynchronously, so they may contain completely different fresh information and have different degrees of staleness, and in addition, the data samples of different clients are likely to be non-identically distributed, which causes data heterogeneity between clients. Second, due to the heterogeneity of computing power, clients with smaller data and better computing power are relatively active and participate in global aggregation more frequently, which may have a greater impact on model aggregation, while clients with larger data and poorer computing power may be ignored, resulting in a decline in the performance of the global model.

[0004] However, at present, most existing asynchronous federated learning methods use static weight aggregation strategies and do not fully consider the dynamic contribution differences and staleness effects between clients. SUMMARY

[0005] Therefore, the application provides an asynchronous federated learning method, device, system and storage medium based on dynamic contribution awareness and staleness adaptation, which can significantly improve the performance of asynchronous federated learning in complex network environments and data distribution while ensuring model accuracy.

[0006] The first object of the application is to provide an asynchronous federated learning method based on dynamic contribution awareness and staleness adaptation.

[0007] The second object of the present application is to provide an asynchronous federated learning device based on dynamic contribution awareness and staleness adaptation.

[0008] The third object of the present application is to provide an asynchronous federated learning system based on dynamic contribution awareness and staleness adaptation.

[0009] The fourth object of the present application is to provide a computer-readable storage medium.

[0010] The first object of the present application can be achieved by adopting the following technical solutions:

[0011] An asynchronous federated learning method based on dynamic contribution awareness and staleness adaptation, applied to a client, the method comprising:

[0012] After the server broadcasts the latest global model and the contribution degree vector, the corresponding contribution degree is received;

[0013] It is judged whether the received contribution degree meets the training condition, if yes, the latest global model is downloaded to train the local model, if not, the old version global model is continued to be used to train the local model;

[0014] The loss function, the model deviation penalty and the staleness inhibition term are fused to construct a local optimization objective function;

[0015] The step size is adjusted according to the historical average delay of the client, and a dynamic learning rate is calculated;

[0016] According to the local optimization objective function and the dynamic learning rate, the gradient descent is executed to update the local model;

[0017] The updated local model is sent to the server, so that the server, after receiving the local models of more than or equal to a preset number of clients, updates the global model by using the dynamic aggregation weight to weight and aggregate the local models, and updates the contribution degree of the client according to the participation state of the client, and generates a contribution degree vector by normalizing the updated contribution degree.

[0018] Further, the judgment of whether the received contribution degree meets the training condition specifically comprises:

[0019] The received contribution degree is compared with the local data proportion, if the received contribution degree is less than the local data proportion, the training condition is met, otherwise, the training condition is not met.

[0020] Further, the loss function, the model deviation penalty and the staleness inhibition term are fused to construct a local optimization objective function, as follows:

[0021]

[0022] Wherein, fk (w k ) is a loss function of the client k, w k is a local model of the client k, is a regularization term, λ is a regularization coefficient, γ t-τ ||w k || 2 is an obsolescence inhibition term, τ is the round when the client k last participated in aggregating the global model, γ is a decay coefficient, is the latest global model.

[0023] Further, the step of adjusting the step size according to the historical average delay of the client, and calculating the dynamic learning rate, specifically includes:

[0024] Adjusting the step size according to the historical average delay of the client, as follows:

[0025]

[0026] wherein, is a dynamic step size factor of the client k, is the historical average delay of the client k, is the delay of the client k at the τth round;

[0027] Calculating the dynamic learning rate using the dynamic step size factor, as follows:

[0028]

[0029] wherein, is a dynamic learning rate of the client k, is a global learning rate.

[0030] Further, the step of performing gradient descent to update the local model according to the local optimization objective function and the dynamic learning rate, as follows:

[0031]

[0032] wherein, is the updated local model of the client k, f k (w k ) is a loss function of the client k, is the current local model of the client k, is a dynamic learning rate of the client k, τ is the round when the client k last participated in aggregating the global model, is the latest global model.

[0033] Further, the step of performing weighted aggregation on the local model using the dynamic aggregation weight, and updating the global model, specifically includes:

[0034] Trigger the aggregation operation, combine the local data proportion of the client, and the current normalized contribution degree of the client to generate a dynamic aggregation weight;

[0035] Using the dynamic aggregation weight, the current local model of the client is weighted and aggregated to update the global model.

[0036] Further, the client contribution degree is updated according to the client participation state, as follows:

[0037]

[0038] Wherein, The current contribution degree of the client k is, The updated contribution degree is β k The local data proportion of the client k is ρ, the global forgetting factor is τ, the round when the client k last participated in the aggregation of the global model is τ, and the decay coefficient is γ.

[0039] The second object of the application can be achieved by adopting the following technical solutions:

[0040] An asynchronous federated learning device based on dynamic contribution perception and obsolescence adaptation, the device comprises:

[0041] The receiving module is configured to receive the corresponding contribution degree after the server broadcasts the latest global model and the contribution degree vector;

[0042] The judging module is configured to judge whether the received contribution degree meets the training condition, if yes, the latest global model is downloaded to train the local model, and if not, the old version of the global model is used to train the local model;

[0043] The construction module is configured to fuse the loss function, the model deviation penalty and the obsolescence inhibition term to construct a local optimization objective function;

[0044] The calculation module is configured to adjust the step size according to the historical average delay of the client and calculate a dynamic learning rate;

[0045] The first updating module is configured to update the local model by gradient descent according to the local optimization objective function and the dynamic learning rate;

[0046] The second updating module is configured to send the updated local model to the server, so that the server updates the global model by weighted aggregation of the local model using the dynamic aggregation weight after receiving the local models of more than or equal to a preset number of clients, and updates the contribution degree of the client according to the client participation state to generate a contribution degree vector by normalizing the updated contribution degree.

[0047] The third object of the application can be achieved by adopting the following technical solutions:

[0048] An asynchronous federated learning system based on dynamic contribution perception and staleness adaptation, the system comprising N clients and a server, the server being connected with the N clients respectively;

[0049] The client is used for executing the asynchronous federated learning method described above.

[0050] The server is used for initializing and broadcasting a global model and a contribution vector, after receiving local models of more than or equal to a preset number of clients, weighting and aggregating the local models by using a dynamic aggregation weight, updating the global model, updating the contribution of the client, and broadcasting the latest global model and the contribution vector.

[0051] The fourth object of the application can be achieved by adopting the following technical solutions:

[0052] A computer-readable storage medium stores a program, the program is executed by a processor to implement the asynchronous federated learning method described above.

[0053] The application has the following beneficial effects relative to the prior art:

[0054] 1. The application significantly improves the communication efficiency and model accuracy of asynchronous federated learning through dynamic contribution perception and staleness adaptation mechanism.

[0055] 2. The application fully considers the dynamic contribution difference and staleness influence between clients through the calculation of the dynamic aggregation weight, thereby improving the performance of the global model.

[0056] 3. The application enhances the robustness of the model to non-independent and identically distributed data through the regularization term and dynamic step adjustment, and has stronger adaptability.

[0057] 4. The application avoids the delay caused by waiting for all clients through the setting of the dynamic aggregation threshold, thereby improving the training efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0058] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description only show some embodiments of the application, and those skilled in the art can obtain other drawings according to the structures shown in these drawings without any creative effort.

[0059] Figure 1 The figure shows the architecture of the asynchronous federated learning system based on dynamic contribution perception and staleness adaptation of the embodiment 1 of the application.

[0060] Figure 2A simple flowchart of the asynchronous federated learning method based on dynamic contribution awareness and staleness adaptation of the embodiment 1 of the present application.

[0061] Figure 3 A detailed flowchart of the asynchronous federated learning method based on dynamic contribution awareness and staleness adaptation of the embodiment 1 of the present application.

[0062] Figure 4 A structural block diagram of the asynchronous federated learning device based on dynamic contribution awareness and staleness adaptation of the embodiment 2 of the present application. DETAILED DESCRIPTION

[0063] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0064] Embodiment 1

[0065] As shown in Figure 1 , the present embodiment provides a personalized federated learning system based on adaptive prototype aggregation, which comprises N clients and a server S. The server S is a central server, which is connected with the N clients respectively. Each client has local data D. Each client also has a neural network model, which is used as a local model.

[0066] As shown in Figure 2 and Figure 3 , the present embodiment also provides an asynchronous federated learning method based on dynamic contribution awareness and staleness adaptation. The method is mainly implemented by the above-mentioned client, and specifically comprises the following steps:

[0067] S301, after the server broadcasts the latest global model and the contribution degree vector, the corresponding contribution degree is received.

[0068] In the present embodiment, the server first performs initialization and parameter setting, specifically: initializing the global model The server initializes the contribution degree vector K represents the number of all clients, the initial value is set to the proportion of the local data of the client The contribution degree decay factor γ, the global forgetting factor ρ and the dynamic aggregation threshold N are set th , wherein N th is a preset number, and when the server receives N thThe local model of the client starts to conduct global aggregation.

[0069] The server broadcasts the latest global model and the contribution degree vector S t It is easy to understand that the latest global model and the contribution degree vector broadcast by the server in the first round are and S 0 The latest global model and the contribution degree vector broadcast in the t-th round are and S t .

[0070] S302, judge whether the received contribution degree meets the training condition.

[0071] In this embodiment, the contribution degree received by the client k in the t-th round of broadcast is The contribution degree is compared with the local data proportion β k If the contribution degree is less than the local data proportion β k , the training condition is met, and the latest global model is downloaded for local training, otherwise, that is, the contribution degree is greater than or equal to the local data proportion β k , the training condition is not met, and the old version of the global model is continued to be used for training to avoid repeated calculation and reduce communication overhead.

[0072] The following steps S303-S305 are the updating process of the local model:

[0073] S303, fuse the loss function, the model deviation penalty and the obsolescence suppression term to construct the local optimization objective function.

[0074] In this embodiment, the client k fuses the loss function, the model deviation penalty and the obsolescence suppression term based on the local data D k to construct the local optimization objective function, as follows:

[0075]

[0076] Wherein, f k (w k ) is the loss function of the client k, such as cross-entropy loss function, w k is the local model of the client k, is the regularization term, λ is the regularization coefficient, for example, λ=0.1, which is used to limit the deviation of the local model from the global model, γ t-τ ||w k || 2 is the obsolescence suppression term, τ is the round number when the client k last participated in the aggregation of the global model, and γ is the attenuation coefficient.

[0077] S304, adjust the step size according to the historical average delay of the client to calculate the dynamic learning rate.

[0078] In this embodiment, the client k first adjusts the dynamic step size, as follows:

[0079]

[0080] wherein, is the dynamic step size factor of the client k, that is, the step size factor of the tth round of broadcast, is the historical average delay of the client k, is the delay of the client k in the th round, and the learning rate amplification coefficient of the client with higher delay is larger.

[0081] Then, the client k calculates the dynamic learning rate using the dynamic step size factor, as follows:

[0082]

[0083] wherein, is the dynamic learning rate of the client k, that is, the learning rate of the tth round of broadcast, is the global learning rate.

[0084] S305, perform gradient descent to update the local model according to the local optimization objective function and the dynamic learning rate.

[0085] In this embodiment, the client k performs gradient descent to update the local model according to the local optimization objective function and the dynamic learning rate, as follows:

[0086]

[0087] wherein, is the updated local model of the client k, f k (w k ) is the loss function of the client k, is the current local model of the client k.

[0088] S306, send the updated local model to the server.

[0089] In this embodiment, the client k sends the updated local model to the server, and the server aggregates the local models of the clients using the dynamic aggregation weight after receiving more than or equal to a preset number of local models, that is, receiving m t ≥N th local models, specifically, triggering the aggregation operation, combining the local data proportion β k of the client and the current normalized contribution degree of the client to generate the dynamic aggregation weight The global model is updated by aggregating the current local models of the clients using dynamic aggregation weights; the server updates the contribution degree of the client according to the participation state of the client, as follows:

[0090]

[0091] wherein, is the current contribution degree of the client k, is the updated contribution degree, and p is a global forgetting factor for suppressing the weight monopoly of long-term active clients.

[0092] The server normalizes the updated contribution degree to generate a contribution degree vector, broadcasts the updated global model and the contribution degree vector, and repeats the above steps S301-S306 until the convergence index or the training round is reached.

[0093] In summary, the method of the present application realizes efficient model training under non-independent and identically distributed data environment, adopts a dynamic aggregation strategy to evaluate and adjust the model updates uploaded by the clients in real time, and introduces a client contribution vector to effectively alleviate the problem of outdated models caused by communication delay; at the same time, a dynamic learning rate adjustment mechanism based on historical delay feedback and an obsolescence suppression strategy are introduced to further improve the model convergence speed and overall performance; both communication efficiency and model accuracy are taken into account, and the method is suitable for large-scale distributed machine learning scenarios, has the advantages of strong robustness and good adaptability, and the comparison with the prior art is shown in Table 1 below.

[0094] Table 1 Comparison of the method of the present application with the prior art

[0095]

[0096]

[0097] It should be noted that although the method operations of the above-described embodiments have been described in a particular, sequential order, this is not necessarily to be understood as a strict sequence or order, unless the schedule is expressly called for. The steps can be performed in any order or concurrently as appropriate, or with additional steps not presented in the embodiments. In addition, some steps can be optional. Additionally or alternatively, some steps can be divided into multiple steps and / or some steps can be combined into a single step.

[0098] Embodiment 2:

[0099] As shown in Figure 4 , the present embodiment provides an asynchronous federated learning device based on dynamic contribution awareness and obsolescence adaptation, which comprises a receiving module 401, a judging module 402, a constructing module 403, a calculating module 404, a first updating module 405 and a second updating module 406, and the specific functions of each module are as follows:

[0100] The receiving module 401 is configured to receive the corresponding contribution degree after the server broadcasts the latest global model and the contribution degree vector;

[0101] The judging module 402 is configured to judge whether the received contribution degree meets the training condition, if yes, the latest global model is downloaded to train the local model, if not, the old version of the global model is used to train the local model;

[0102] The constructing module 403 is configured to fuse the loss function, the model deviation penalty and the obsolescence inhibition term to construct a local optimization objective function;

[0103] The computing module 404 is configured to adjust the step size according to the historical average delay of the client to calculate a dynamic learning rate;

[0104] The first updating module 405 is configured to perform gradient descent to update the local model according to the local optimization objective function and the dynamic learning rate;

[0105] The second updating module 406 is configured to send the updated local model to the server, so that the server performs weighted aggregation on the local model by using a dynamic aggregation weight to update the global model and update the contribution degree of the client after receiving more than or equal to a preset number of local models of the clients.

[0106] It should be noted that the device provided in the embodiment is only exemplified by the division of the above-mentioned functional modules, and in actual application, the above-mentioned functions can be completed by different functional modules according to needs, that is, the internal structure is divided into different functional modules to complete all or part of the functions described above.

[0107] Embodiment 3

[0108] The embodiment provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the asynchronous federated learning method of the above-mentioned embodiment 1, as follows:

[0109] After the server broadcasts the latest global model and the contribution degree vector, the corresponding contribution degree is received; it is judged whether the received contribution degree meets the training condition, if yes, the latest global model is downloaded to train the local model, if not, the old version global model is used to train the local model; the loss function, model deviation penalty and obsolescence inhibition term are fused to construct a local optimization objective function; the step size is adjusted according to the historical average delay of the client, and the dynamic learning rate is calculated; the local model is updated by gradient descent according to the local optimization objective function and the dynamic learning rate; the updated local model is sent to the server, so that the server aggregates the local models of the clients by using the dynamic aggregation weight after receiving more than or equal to a preset number of local models, updates the global model, and updates the contribution degree of the client.

[0110] It should be noted that the computer readable storage medium of the embodiment can be a computer readable signal medium or a computer readable storage medium, or any combination of the two. The computer readable storage medium may, for example, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor device, device or device, or any combination of the above. More specific examples of computer readable storage medium can include, but are not limited to: electrical connection with one or more conductors, portable computer disk, hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above.

[0111] In the embodiment, the computer readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution device, device or device. In the embodiment, the computer readable signal medium can include a data signal propagating in a baseband or as a part of a carrier wave, which carries a computer readable program. Such a propagating data signal can take many forms, including but not limited to electromagnetic signals, optical signals or any suitable combination of the above. The computer readable signal medium can also be any computer readable storage medium other than the computer readable storage medium, which can send, propagate or transmit a program for use by or in conjunction with an instruction execution device, device or device. The computer program contained on the computer readable storage medium can be transmitted by any suitable medium, including but not limited to: wire, cable, RF (radio frequency), etc., or any suitable combination of the above.

[0112] The computer readable storage medium described above can be used to store computer programs for executing the embodiments of the present application, which can be written in one or more programming languages or combinations thereof, including object oriented programming languages such as Java, Python, C++ and conventional procedural programming languages such as C programming language or similar programming languages. The program can be executed entirely on the user computer, partially on the user computer, as an independent software package, partially on the user computer and partially on a remote computer, or entirely on a remote computer or server. In the case involving a remote computer, the remote computer can be connected to the user computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (for example, through the Internet by using an Internet service provider).

[0113] The above description is merely illustrative of the embodiments of the present application, and the present application is not limited to the details of the above exemplary embodiments. It will be apparent to those skilled in the art that the present application can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. Therefore, the embodiments should be considered in a descriptive sense only and not for purposes of limitation. The scope of the present application should be defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the claims are intended to be embraced therein.

Claims

1. An asynchronous federated learning method based on dynamic contribution awareness and staleness adaptation, applied to the client side, characterized in that, The method includes: After the server broadcasts the latest global model and contribution vector, the corresponding contribution is received. Determine whether the received contribution meets the training conditions. If yes, download the latest global model to train the local model. If not, continue to use the old version of the global model to train the local model. By fusing the loss function, model deviation penalty, and staleness suppression term, a local optimization objective function is constructed. The step size is adjusted based on the client's historical average latency, and the dynamic learning rate is calculated. Based on the local optimization objective function and dynamic learning rate, perform gradient descent to update the local model; The updated local model is sent to the server so that after the server receives more than or equal to a preset number of client local models, it uses dynamic aggregation weights to perform weighted aggregation of the local models, updates the global model, updates the client contribution based on the client participation status, and normalizes the updated contribution to generate a contribution vector. The fusion loss function, model deviation penalty, and staleness suppression term are used to construct the local optimization objective function, as shown in the following equation: Among them, f k (w k Let w be the loss function for client k. k For the local model of client k, γ is the regularization term, λ is the regularization coefficient, and γ is the regularization term. t-τ ||w k || 2 The term represents the staleness suppression term, where τ is the round in which client k most recently participated in the aggregation of the global model, and γ is the decay coefficient. This is the latest global model; The step size adjustment based on the historical average latency of the client, and the calculation of the dynamic learning rate, specifically include: The step size is adjusted based on the client's historical average latency, as follows: in, The dynamic step size factor for client k. Let k be the historical average latency of client k. Let be the delay for client k in round τ; The dynamic learning rate is calculated using the dynamic step size factor, as follows: in, The dynamic learning rate for client k. The global learning rate; The local model is updated by performing gradient descent based on the local optimization objective function and dynamic learning rate, as shown in the following equation: in, For the local model updated for client k, f k (w k Let be the loss function for client k. For the current local model of client k, Let τ be the dynamic learning rate of client k, and τ be the round in which client k last participated in aggregating the global model. This is the latest global model.

2. The asynchronous federated learning method according to claim 1, characterized in that, The determination of whether the received contribution meets the training conditions specifically includes: The received contribution is compared with the proportion of local data. If the received contribution is less than the proportion of local data, the training condition is met; otherwise, the training condition is not met.

3. The asynchronous federated learning method according to claim 1, characterized in that, The step of using dynamic aggregation weights to perform weighted aggregation on the local model and updating the global model specifically includes: Trigger the aggregation operation and generate dynamic aggregation weights by combining the proportion of local data on the client and the client's current normalized contribution. By using dynamic aggregation weights, the current local model on the client is weighted and aggregated to update the global model.

4. The asynchronous federated learning method according to claim 1, characterized in that, The method for updating the client contribution based on the client's participation status is as follows: in, For client k, the current contribution level. For the updated contribution, β k Let ρ be the proportion of local data for client k, ρ be the global forgetting factor, τ be the round in which client k last participated in the aggregation of the global model, and γ be the decay coefficient.

5. An asynchronous federated learning device based on dynamic contribution perception and staleness adaptation, characterized in that, The device includes: The receiving module is used to receive the corresponding contribution after the server broadcasts the latest global model and contribution vector; The judgment module determines whether the received contribution meets the training conditions. If so, it downloads the latest global model to train the local model; otherwise, it continues to use the old version of the global model to train the local model. The building module is used to fuse the loss function, model deviation penalty, and staleness suppression term to construct a local optimization objective function; The calculation module is used to adjust the step size based on the client's historical average latency and calculate the dynamic learning rate. The first update module is used to perform gradient descent to update the local model based on the local optimization objective function and dynamic learning rate; The second update module is used to send the updated local model to the server so that after the server receives more than or equal to a preset number of client local models, it uses dynamic aggregation weights to perform weighted aggregation on the local models, updates the global model, updates the client contribution based on the client participation status, and normalizes the updated contribution to generate a contribution vector. The fusion loss function, model deviation penalty, and staleness suppression term are used to construct the local optimization objective function, as shown in the following equation: Among them, f k (w k Let w be the loss function for client k. k For the local model of client k, γ is the regularization term, λ is the regularization coefficient, and γ is the regularization term. t-τ ||w k || 2 The term represents the staleness suppression term, where τ is the round in which client k most recently participated in the aggregation of the global model, and γ is the decay coefficient. This is the latest global model; The step size adjustment based on the historical average latency of the client, and the calculation of the dynamic learning rate, specifically include: The step size is adjusted based on the client's historical average latency, as follows: in, The dynamic step size factor for client k. Let k be the historical average latency of client k. Let be the delay for client k in round τ; The dynamic learning rate is calculated using the dynamic step size factor, as follows: in, Let η be the dynamic learning rate for client k, and η be the global learning rate. The local model is updated by performing gradient descent based on the local optimization objective function and dynamic learning rate, as shown in the following equation: in, For the local model updated for client k, f k (w k Let be the loss function for client k. For the current local model of client k, Let τ be the dynamic learning rate of client k, and τ be the round in which client k last participated in aggregating the global model. This is the latest global model.

6. An asynchronous federated learning system based on dynamic contribution awareness and staleness adaptation, characterized in that, The system includes N clients and a server, with each server connected to one of the N clients. The client is used to execute the asynchronous federated learning method according to any one of claims 1-4; The server is used to initialize the global model and contribution vector and broadcast them. After receiving a number of local client models greater than or equal to a preset number, it uses dynamic aggregation weights to perform weighted aggregation on the local models, updates the global model, updates the client contribution, and broadcasts the latest global model and contribution vector.

7. A computer-readable storage medium storing a program, characterized in that, When the program is executed by the processor, it implements the asynchronous federated learning method according to any one of claims 1-4.

Images