A quality-aware online client selection method and apparatus

Abstract

The application discloses a quality-aware based online client selection method and device, calculates the data quality of the tth round of clients, then calculates the communication delay reward estimate and the communication energy consumption reward estimate according to the cumulative number of times that the t-1th round of users select channels, allocates channels to each client, calculates the training delay reward estimate and the training energy consumption reward estimate according to the cumulative number of times that the t-1th round of servers select clients, calculates the appropriate client set through the communication delay reward estimate, the communication energy consumption reward estimate, the training delay reward estimate, the training energy consumption reward estimate and the data quality, and finally recalculates the channel allocation for the selected client set until all channel allocations are completed, so as to realize the quality-aware based online client selection. Through the design of the joint optimization strategy such as the client selection and the channel allocation, better performance can be shown, the balance between the data quality, the delay and the energy consumption is realized, and the convergence precision is improved.

Description

Technical Field

[0001] This application relates to the field of communication technology, and more specifically, to a method and apparatus for selecting online clients based on quality awareness. Background Technology

[0002] Traditional machine learning requires collecting large amounts of data and uploading it to servers for model training, resulting in expensive communication costs. As users become more aware of data security, clients are reluctant to actively share local data, leading to fragmented data ownership and data silos. To address these issues, Federated Learning (FL) was proposed: clients train models locally and send them to a server for model aggregation. The server then sends the aggregated model back to the client for iterative training until convergence. However, due to system heterogeneity, statistical heterogeneity, malicious attacks, and the uncertainty of dynamic environments, the training process of federated learning often requires a long time to converge or may not converge at all.

[0003] However, the contribution of clients to training performance is closely related to their computing power, communication capabilities, and vulnerability. Therefore, selecting appropriate clients is essential for efficient federated learning. However, most work focuses on latency and energy consumption, neglecting the impact of local dataset quality on the training process. Furthermore, due to variations in channel conditions and available computing resources, the server needs to select suitable clients for training based on real-time communication and computing resources.

[0004] Based on existing research, two issues remain unresolved. Firstly, data quality, representing the distribution of client datasets and their susceptibility to malicious attacks, is crucial for convergence. Currently, there is no comprehensive metric to describe data quality. Therefore, how to assess data quality urgently needs to be addressed. Secondly, much of the research on client scheduling focuses on selecting clients with low latency, low energy consumption, or high data quality. Considering only one aspect to ensure training effectiveness is meaningless. Considering the balance between data quality, latency, and energy consumption is highly practical. Summary of the Invention

[0005] The purpose of this application is to overcome the shortcomings of existing technologies and provide a quality-aware online client selection method. By designing joint optimization strategies for client selection and channel allocation, this method can exhibit better performance, achieve a balance between data quality, latency, and energy consumption, and improve convergence accuracy.

[0006] The objective of this application is achieved through the following technical solution:

[0007] Firstly, this application proposes a quality-aware online client selection method, applied to a wireless federated learning system, the wireless federated learning system comprising a server and multiple clients connected to the server, the method comprising:

[0008] Calculate the data quality of the client in round t, where t is the time slot cycle variable;

[0009] The communication latency reward estimate and the communication energy consumption reward estimate are calculated based on the cumulative number of times the user selects a channel in round t-1. The communication latency reward estimate and the communication energy consumption reward estimate are used to allocate a channel to each client.

[0010] The training latency reward estimate and training energy consumption reward estimate are calculated based on the cumulative number of times the server selects the client in round t-1.

[0011] The communication latency reward estimate, the communication energy consumption reward estimate, the training latency reward estimate, the training energy consumption reward estimate, and the data quality calculation client set are used to allocate channels to users.

[0012] The channel allocation is recalculated for the client set until all channels are allocated, in order to achieve quality-aware online client selection.

[0013] In one possible implementation, prior to the step of calculating the data quality of the client in round t, the method further includes:

[0014] Initialize client data and server;

[0015] The client data includes settings for time slot length, client, number of channels, time slot loop variable, client loop variable, and channel loop variable.

[0016] In one possible implementation, the step of calculating the data quality of the client in round t includes:

[0017] Calculate the heterogeneity of the client in round t based on the client's local training loss and testing loss;

[0018] Calculate the attack level of the client in round t based on the client's global loss and test loss;

[0019] The client data quality in round t is calculated by combining the heterogeneity and attack vulnerability of the client in round t.

[0020] In one possible implementation, the client data quality v in the t-th round k The formula for calculating (t) is:

[0021]

[0022] Where x k (-1) indicates whether the client is selected to participate in the training process in round t-1, v k (-1) represents the client data quality in round t-1, a k (t) represents the heterogeneity of the client in round t, b k (t) represents the attack level of the client in round t, α1 represents the heterogeneity weight factor of the client in round t, and α2 is the attack level weight factor of the client in round t.

[0023] In one possible implementation, the communication latency reward valuation The calculation formula is:

[0024]

[0025] in, Let z be the average communication delay reward for the client to select a channel in round t-1. k,m (t-1) is the first cumulative number of times the user selects a channel in round t-1;

[0026] The estimated communication energy consumption reward The calculation formula is:

[0027]

[0028] in, Let z be the average communication energy consumption reward for the client to select a channel in round t-1. k,m (t-1) is the first cumulative number of times the user selects a channel in the (t-1)th round.

[0029] In one possible implementation, the training latency reward estimation The calculation formula is:

[0030]

[0031] in, Let y be the average training latency reward for the client during round t-1. k (t-1) is the cumulative number of times the server selects a client in the (t-1)th round;

[0032] The training energy consumption reward valuation The calculation formula is:

[0033]

[0034] in, Let y be the average training energy consumption reward for the client in round t-1. k(t-1) is the cumulative number of times the server selects a client in round t-1.

[0035] In one possible implementation, the step of allocating a channel to each client is as follows:

[0036] Let two vectors A(k) and B be defined. k (m) represents the sum of the estimated communication latency reward and the estimated communication energy consumption reward;

[0037] Let vector A(k) be sorted in descending order, and calculate the performance ranking of the channel among clients.

[0038] Let vector B(m) be sorted in descending order, and calculate the performance ranking of the client in the channel.

[0039] Channels are assigned to clients based on their performance ranking among clients. It is defined whether a client selects a channel for data transmission in round t. If the value is 1, the client selects a channel for data transmission; otherwise, the client does not select a channel.

[0040] If multiple clients select the same channel, the channel is selected based on the client's performance ranking within the channel. This process is repeated to select the next channel for other clients based on their performance ranking within the channel.

[0041] Secondly, this application also proposes a quality-aware online client selection device, the device comprising:

[0042] The quality calculation module is used to calculate the data quality of the client in round t, where t is the time slot cycle variable;

[0043] The communication valuation calculation module is used to calculate the communication latency reward valuation and the communication energy consumption reward valuation based on the cumulative number of times the user selects a channel in the (t-1)th round. The communication latency reward valuation and the communication energy consumption reward valuation are used to allocate a channel to each client.

[0044] The training valuation module is used to calculate the training latency reward valuation and the training energy consumption reward valuation based on the cumulative number of times the server selects the client in round t-1.

[0045] The client set calculation module is used to calculate the client set using the communication latency reward estimate, the communication energy consumption reward estimate, the training latency reward estimate, the training energy consumption reward estimate, and the data quality, wherein the client set allocates channels to users;

[0046] The allocation module is used to recalculate channel allocations for the client set until all channels are allocated, in order to achieve quality-aware online client selection.

[0047] Thirdly, this application also proposes a computer device comprising a processor and a memory, wherein the memory stores a computer program that is loaded and executed by the processor to implement a quality-aware online client selection method as described in any of the first aspects.

[0048] Fourthly, this application also proposes a computer-readable storage medium storing a computer program that is loaded and executed by a processor to implement a quality-aware online client selection method as described in any of the first aspects.

[0049] The main solution and its various further alternatives described above can be freely combined to form multiple solutions, all of which are solutions that can be adopted and are claimed in this application; furthermore, the (non-conflicting alternatives) can also be freely combined with each other and with other alternatives. Those skilled in the art, after understanding the solution of this application, will realize from the prior art and common general knowledge that there are many combinations, all of which are technical solutions to be protected by this application, and will not be exhaustively listed here.

[0050] The beneficial effects of this application are as follows:

[0051] This application discloses a quality-aware online client selection method. On the one hand, it constructs a data quality index to evaluate the heterogeneity of the local dataset and the degree of label-flipping attack, and designs a joint optimization strategy including client selection and channel allocation. On the other hand, it develops an efficient algorithm that does not require prior knowledge to solve the optimization variables, thereby achieving a balance between data quality, latency and energy consumption and improving convergence accuracy. Attached Figure Description

[0052] Figure 1 A flowchart illustrating the online client selection method based on quality awareness proposed in this application is shown.

[0053] Figure 2 The diagram illustrates the relationship between global accuracy and the number of training epochs under different algorithms.

[0054] Figure 3 The diagram illustrates the relationship between global accuracy and clock time under different algorithms.

[0055] Figure 4 The diagram illustrates the relationship between the ratio of selected attack users and the proportion of attack users under different algorithms.

[0056] Figure 5 The diagram illustrates the relationship between the cumulative number of failed clients and the number of training rounds under different algorithms.

[0057] Figure 6 The diagram illustrates the relationship between total reward and number of training rounds under different algorithms.

[0058] Figure 7 The diagram illustrates the relationship between average device data quality and the number of training rounds under different algorithms.

[0059] Figure 8 The diagram illustrates the relationship between average device latency and the number of training rounds under different algorithms.

[0060] Figure 9 The diagram illustrates the relationship between average device power consumption and the number of training rounds under different algorithms. Detailed Implementation

[0061] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0062] Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0063] Two problems remain unresolved in current technology. Firstly, data quality, representing the distribution of client datasets and whether they are affected by malicious attacks, is crucial for convergence. Currently, there is no comprehensive metric to describe data quality. Therefore, how to assess data quality is a pressing issue. Secondly, much of the research on client scheduling focuses on selecting clients with low latency, low energy consumption, or high data quality. Considering only one aspect to ensure training effectiveness is meaningless. Considering the balance between data quality, latency, and energy consumption is highly practical.

[0064] Therefore, in order to solve the above problems, this application proposes a quality-aware online client selection method. By designing joint optimization strategies such as client selection and channel allocation, it can exhibit better performance, achieve a balance between data quality, latency and energy consumption, and improve convergence accuracy. The following is a detailed description of the method.

[0065] Please refer to Figure 1 , Figure 1The diagram illustrates a flowchart of a quality-aware online client selection method proposed in an embodiment of this application. This method is applied to a wireless federated learning system, which includes a server and multiple clients connected to the server. The method includes the following steps:

[0066] S100, Calculate the data quality of the client in round t.

[0067] S200. Calculate the communication delay reward estimate and the communication energy consumption reward estimate based on the cumulative number of times the user selects a channel in round t-1.

[0068] S300: Calculate the training latency reward estimate and training energy consumption reward estimate based on the cumulative number of times the server selects the client in round t-1.

[0069] S400 calculates client sets through communication latency reward estimation, communication energy consumption reward estimation, training latency reward estimation, training energy consumption reward estimation, and data quality calculation.

[0070] S500: Recalculate channel allocation for the client set until all channels are allocated, in order to achieve quality-aware online client selection.

[0071] Where t is the time slot cyclic variable, the communication delay reward estimate and the communication energy consumption reward estimate are used to allocate channels for each client, and the client set is used to allocate channels for users.

[0072] First, starting from round t = K+1, calculate and update the data quality v of client k in round t. k (t), and then the cumulative number of times user k selects channel m in the (t-1)th round z. k,m (t-1), Mean of communication delay reward and the average communication energy consumption bonus Calculate and update the estimated communication latency reward for round t. And communication energy consumption incentive valuation Based on communication latency reward valuation And communication energy consumption incentive valuation Pre-allocate a channel for each client k Then, the cumulative number y of client k selected from the server in round t-1 is... k (t-1), Mean of training latency reward and the average training energy reward Calculate and update the training latency reward estimate for round t. and training energy reward valuation

[0073] Based on communication latency reward valuation Communication energy consumption incentive valuation Training latency reward valuation Training energy reward valuation And the data quality v of client k k (t) Calculate the set of clients that meet the requirements in round t. And re-set for the client The user is assigned a channel. The cumulative number of times client k selects channel m in round t is last updated z. k,m (t), Mean of communication latency reward Average energy consumption bonus for communication The cumulative number of times the server selects client k (y) k (t), Mean of training latency reward training energy reward mean Continuous iterative updates are implemented to achieve quality-aware online client selection.

[0074] In one possible implementation, prior to the step of calculating the data quality of the client in round t, the method further includes:

[0075] Initialize client data and server.

[0076] Client data includes settings for time slot length, client, number of channels, time slot loop variable, client loop variable, and channel loop variable.

[0077] Since the wireless federated learning system consists of a server and multiple clients connected to the server, and these clients can be evenly distributed within an area of ​​500m radius, with the server located at the center of this area and having M channels between it and the clients, and assuming the system has T rounds of global communication, the client set is defined as follows. The channel set is The set of rounds is Initialize client data and server, and set the loop variable for the number of rounds. Set loop variable client Set the loop variable channel For the first K training rounds, only one client is selected to participate in training in each time slot, and all channels are assigned to that client in each round.

[0078] Furthermore, in this embodiment, the reward for the client set can be calculated according to the following optimization problem:

[0079] P1:

[0080]

[0081] stx k (t)∈{0,1},

[0082] ρk,m (t)∈{0,1},

[0083]

[0084]

[0085]

[0086] Where λ1, λ2, and λ3 are positive parameters that adjust the balance between data quality, latency, and energy consumption, respectively, and x k ρ(t) indicates whether client k is selected to participate in the training process in round t-1, where 1 represents participation and no value represents non-participation. k,m (t) indicates whether client k selects channel m for data transmission in round t. A value of 1 indicates that channel m is selected for data transmission; otherwise, it is not selected. k (t) represents the data quality of client k in round t. This represents the training time of client k in round t. This represents the communication time of client k in round t. This represents the training energy consumption of client k in round t. Let τ represent the communication energy consumption of client k in round t. max E is the maximum round time in each round. max It is the maximum energy consumption for each round.

[0087] The step of calculating the client's data quality in round t in step S100 includes:

[0088] Calculate the heterogeneity of the client in round t based on the client's local training loss and testing loss;

[0089] Calculate the attack level of the client in round t based on the client's global loss and test loss;

[0090] The client data quality in round t is calculated by combining the heterogeneity and attack level of the client in round t.

[0091] Among them, data quality v k (t) is jointly evaluated by the heterogeneity and attack vulnerability of the clients. Heterogeneity measures the degree of non-independent distribution of clients, while attack vulnerability measures whether a client is subject to malicious data attacks and the extent of such attacks. The calculation methods for both are the same as those for the selected client set. The training results are related to the test results.

[0092] In one possible implementation, the heterogeneity a of client k in round t is first calculated. k (t), in, It is the local model of client k in the previous round. Training loss, It is the test loss of client k in the previous round, and then the attack degree b of client k in round t is calculated. k (t), Among them, F test ( t-2 ) is the global model w from the previous round on client k. t-2 The test loss below, It represents the test loss of client k in the previous round.

[0093] Client data quality in round t k The formula for calculating (t) is:

[0094]

[0095] Where x k (-1) indicates whether the client is selected to participate in the training process in round t-1, v k (-1) represents the client data quality in round t-1, a k (t) represents the heterogeneity of the client in round t, b k (t) represents the attack level of the client in round t, α1 represents the heterogeneity weight factor of the client in round t, and α2 is the attack level weight factor of the client in round t.

[0096] Since the computational resources and channel conditions of the current round are unknown, it is impossible to obtain the latency and energy consumption of the current round a priori. To solve this problem, embodiments of this application estimate latency and energy consumption by extending the UCB algorithm, utilizing past high-performing behaviors to explore potentially high-return behaviors to maximize future returns.

[0097] In step S200, the estimated communication delay reward is calculated based on the cumulative number of times the user selects a channel in round t-1. The calculation formula is:

[0098]

[0099] in, Let z be the average communication delay reward for the client to select a channel in round t-1. k,m (t-1) is the first cumulative number of times the user selects a channel in round t-1.

[0100] In step S200, China calculates the estimated communication energy consumption reward based on the cumulative number of times the user selects a channel in round t-1. The calculation formula is:

[0101]

[0102] in, Let z be the average communication energy consumption reward for the client to select a channel in round t-1. k,m (t-1) is the first cumulative number of times the user selects a channel in the (t-1)th round.

[0103] When allocating channels to each client based on communication latency reward estimates and communication energy consumption reward estimates, the optimal channel is chosen. Since the computational resources and channel conditions for the current round are unknown, the UCB estimate can be used as a pre-allocation metric. During the channel allocation phase, the client acts as a player, and the channel as an arm. Each player sequentially selects the optimal arm to maximize their reward.

[0104] In one possible implementation, the step of allocating a channel to each client is as follows:

[0105] Let two vectors A(k) and B be defined. k (m) represents the sum of the estimated communication latency reward and the estimated communication energy consumption reward;

[0106] Let vector A(k) be sorted in descending order, and calculate the performance ranking of the channel among clients.

[0107] Let vector B(m) be sorted in descending order, and calculate the performance ranking of the client in the channel.

[0108] Channels are assigned to clients based on their performance ranking among clients. It is defined whether a client selects a channel for data transmission in round t. If the value is 1, the client selects a channel for data transmission; otherwise, the client does not select a channel.

[0109] If multiple clients select the same channel, the channel is selected based on the client's performance ranking within the channel. This process is repeated to select the next channel for other clients based on their performance ranking within the channel.

[0110] First, define two vectors. and Let A(k) be sorted in descending order. Calculate the performance ranking of the M channels at client k. The ranking set is represented as... For client k, the best channel is Let B(m) be sorted in descending order. Calculate the performance ranking of the K clients on channel m, and represent the ranking set as follows: For channel m, the best client is according to Assign the best channel to each client. If multiple clients select the same channel according to For channel Select the best client k, and repeat the previous step according to... Choose the next best channel for other clients; otherwise, let This indicates that client k selects a channel.

[0111] In step S300, the training latency reward estimate is calculated based on the cumulative number of times the server selects a client in round t-1. The calculation formula is:

[0112]

[0113] in, Let y be the average training latency reward for the client during round t-1. k (t-1) is the cumulative number of times the server selects a client in the (t-1)th round.

[0114] In step S300, the training energy consumption reward estimate is calculated based on the cumulative number of times the server selects a client in round t-1. The calculation formula is:

[0115]

[0116] in, Let y be the average training energy consumption reward for the client in round t-1. k (t-1) is the cumulative number of times the server selects a client in round t-1.

[0117] In step S400, the communication delay reward is used for estimation. Communication energy consumption incentive valuation Training latency reward valuation Training energy reward valuation and data quality v k (t) Calculate the set of clients that meet the requirements in round t. In the client selection phase, the server can be considered the player, and the client the arm. Players can choose multiple arms to create a super arm to maximize their rewards. If... Let x k (t) = 1,

[0118] In obtaining client set Then reassemble for the client User-allocated channels are used to ensure efficient resource utilization. The client participates in the training, assigning all channels to Client usage, not The client is not assigned a channel. Channel performance is ranked among the clients. For the selected client set Assign each the best channel If multiple clients select the same channel according to For channel Select the best client k, and repeat the process of selecting the next best channel for the other clients. Otherwise, let This indicates that client k selects a channel. If there are still channels that have not been assigned to clients, repeat the above operation until all channels are assigned.

[0119] Finally, in step S500, the channel allocation for the client set is recalculated until all channels are allocated. The steps to achieve a quality-aware online client are as follows: at the end of each training and communication round, the communication delay of client k in round t transmitting on channel m is observed. Communication energy consumption of client k transmitting on channel m Training delay of client k transmitting on channel m Communication energy consumption of client k transmitting on channel m It is necessary to base it on this The cumulative number of times the client selects channel m for client k. k,m (t), Mean of communication latency reward Average energy consumption bonus for communication The cumulative number of times the server selects client k (y) k (t), Mean of training latency reward training energy reward mean An equivalent update.

[0120] This involves calculating and updating the cumulative number of times client k selects channel m in round t. ρ k,m (t) indicates whether client k selects channel m for data transmission in round t. If it is 1, channel m is selected for data transmission; otherwise, it is not selected.

[0121] Calculate and update the average communication latency reward for round t. It is the communication delay reward for client k transmitting data on channel m. τ is the communication delay of client k transmitting on channel m. max It is the maximum total latency for each round.

[0122] Calculate and update the average communication energy consumption reward for round t. It is the communication energy reward for client k transmitting data on channel m. E is the communication energy consumption of client k transmitting data on channel m in round t. max It is the maximum total energy consumption for each round.

[0123] Calculate and update the cumulative number of times the server selects client k in round t. x k (t) indicates whether client k is selected to participate in the training process in round t. A value of 1 indicates participation, while a value other than 1 indicates non-participation.

[0124] Calculate and update the average latency reward for client k in round t. It is the communication delay reward for client k transmitting data on channel m. τ is the training delay of client k transmitting on channel m. max It is the maximum total latency for each round.

[0125] Calculate and update the average training energy reward. It is the training energy reward for client k transmitting data on channel m. E is the training energy consumption of client k transmitting on channel m in round t. max It is the maximum total energy consumption for each round.

[0126] In one possible implementation, the quality-aware online client selection method proposed in the embodiments of this application is verified. Please refer to... Figure 2 , Figure 2 The diagram illustrates the relationship between global accuracy and the number of training epochs for different algorithms. Random and Fixed are compared algorithms; this embodiment proposes the DQ-UCB algorithm. The percentage of clients subjected to tag attacks was initially set to 0%, 20%, and 40%. In the presence of attacking users, the DQ-UCB algorithm exhibits higher convergence accuracy than the Random and Fixed algorithms. Furthermore, as the number of attacked clients increases, the DQ-UCB algorithm is less affected, while the convergence accuracy of the other two algorithms drops sharply.

[0127] Please refer to Figure 3 , Figure 3 The diagram illustrates the relationship between global accuracy and clock time under different algorithms. Regardless of the specific algorithm, the DQ-UCB algorithm achieves convergence in less time than the Random and Fixed algorithms. Figure 4The diagram illustrates the relationship between the percentage of users selected for attack and the percentage of attacking users under different algorithms. When attacking users account for 20%, the DQ-UCB algorithm selects attacking users only 1.35% of the time; when attacking users account for 40%, the percentage is only 3.25%. However, the other two algorithms consistently select attacking users at a rate close to 50% in all cases. Figure 2 and Figure 4 The DQ-UCB algorithm in this application can discard unreliable clients and select better clients to participate in training.

[0128] Please refer to Figure 5 , Figure 5 The diagram illustrates the relationship between the cumulative number of failed clients and the number of training rounds under different algorithms. If a client in S(t) fails to meet the constraints, it cannot send an update to the server for aggregation. Figure 5 It can be seen that the algorithm proposed in this invention has almost no failed clients, while the other two algorithms have a sharp increase in failed clients. This indicates that the algorithm proposed in this invention can discard clients with poor computing and communication capabilities to achieve latency and energy consumption constraints.

[0129] Figure 6 The diagram illustrates the relationship between total reward and number of training rounds for different algorithms. The total reward of the DQ-UCB algorithm in this application is consistently higher than that of the Random and Fixed algorithms. Figure 7 The diagram illustrates the relationship between average device data quality and training rounds under different algorithms. The time-averaged data quality of the DQ-UCB algorithm in this application is consistently higher than that of the Random and Fixed algorithms. Figure 8 A schematic diagram illustrating the relationship between average device latency and training rounds for different algorithms is shown. The average latency of the DQ-UCB algorithm in this application is consistently lower than that of the Random and Fixed algorithms. Finally... Figure 9 A schematic diagram illustrating the relationship between average device power consumption and training rounds for different algorithms is shown. The time-averaged power consumption of the DQ-UCB algorithm in this application is consistently lower than that of the Random and Fixed algorithms. Combined with... Figure 7 , Figure 8 and Figure 9 The DQ-UCB algorithm in this application can discard clients with poor computing and communication capabilities to achieve the purpose of time and energy consumption constraints.

[0130] The simulation results above demonstrate that the quality-aware online client selection method proposed in this application can achieve efficient resource allocation and maximize average total reward. The selected client set can discard clients with poor data quality, poor channel conditions, and poor computing power, resulting in good convergence of the algorithm. Furthermore, the joint optimization scheme in this algorithm exhibits better performance than the random scheme in terms of average total reward, average latency, average energy consumption, and average data quality.

[0131] Furthermore, this application proposes a quality-aware online client selection framework for wireless federated learning systems. Within this framework, a data quality metric is designed to evaluate the heterogeneity of local datasets and their vulnerability to label-flipping attacks. A reward model is constructed based on a linearly weighted combination of multiple factors, including data quality, latency, and energy consumption. An efficient algorithm is developed that requires no CPU computing resources or prior knowledge of wireless CSI.

[0132] Compared with the prior art, the embodiments of this application have the following beneficial effects:

[0133] First, we constructed a data quality metric to assess the heterogeneity of local datasets and their vulnerability to label flipping attacks, and designed a joint optimization strategy that includes client selection and channel allocation.

[0134] Second, the joint optimization scheme proposed in this application shows better performance than the stochastic method in terms of average total reward, average data quality, average latency and average energy consumption, and can achieve higher convergence accuracy and converge in less time.

[0135] Furthermore, this application also proposes a quality-aware online client selection device, which includes:

[0136] The quality calculation module is used to calculate the data quality of the client in round t, where t is the time slot cycle variable;

[0137] The communication valuation calculation module is used to calculate the communication latency reward valuation and the communication energy consumption reward valuation based on the cumulative number of times the user selects a channel in round t-1. The communication latency reward valuation and the communication energy consumption reward valuation are used to allocate a channel to each client.

[0138] The training valuation module is used to calculate the training latency reward valuation and the training energy consumption reward valuation based on the cumulative number of times the server selects the client in round t-1.

[0139] The client set calculation module is used to calculate the client set based on communication latency reward estimate, communication energy consumption reward estimate, training latency reward estimate, training energy consumption reward estimate, and data quality. The client set is then used to allocate channels to users.

[0140] The allocation module is used to recalculate channel allocations for the client set until all channels are allocated, in order to enable quality-aware online client selection.

[0141] This preferred embodiment provides a computer device that can implement the steps of any embodiment of the quality-aware online client selection method provided in this application. Therefore, it can achieve the beneficial effects of the quality-aware online client selection method provided in this application, as detailed in the preceding embodiments, which will not be repeated here.

[0142] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor. Therefore, embodiments of this application provide a storage medium storing multiple instructions that can be loaded by a processor to execute the steps of any embodiment of the quality-aware online client selection method provided in this application.

[0143] The storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0144] Since the instructions stored in the storage medium can execute the steps in any of the quality-aware online client selection method embodiments provided in this application, the beneficial effects that any of the quality-aware online client selection methods provided in this application can achieve can be realized. For details, please refer to the previous embodiments, which will not be repeated here.

[0145] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A quality-aware online client selection method, characterized in that, The method is applied to a wireless federated learning system, which includes a server and multiple clients connected to the server. The method includes: Calculate the data quality of the client in round t, where t is the time slot cycle variable; The communication latency reward estimate and the communication energy consumption reward estimate are calculated based on the cumulative number of times the user selects a channel in round t-1. The communication latency reward estimate and the communication energy consumption reward estimate are used to allocate a channel to each client. The steps for allocating a channel to each client are as follows: Set two vectors and Both are the sum of the estimated communication latency reward and the estimated communication energy consumption reward; Let vector Sort in descending order and rank the channel performance among clients. , ; Let vector Sort clients in descending order and rank them by channel performance. , ; Channels are assigned to clients based on their performance ranking within the client, as defined in... The system prompts the client to select a channel for data transmission. If the value is 1, the client selects a channel for data transmission; otherwise, it does not. If multiple clients select the same channel, the channel is selected based on the client's performance ranking within the channel. If the operation is repeated, the next channel is selected for other clients based on the channel's performance ranking among the clients. The training latency reward estimate and training energy consumption reward estimate are calculated based on the cumulative number of times the server selects the client in round t-1. The communication latency reward estimate, the communication energy consumption reward estimate, the training latency reward estimate, the training energy consumption reward estimate, and the data quality calculation client set are used to allocate channels to users. The channel allocation is recalculated for the client set until all channels are allocated, in order to achieve quality-aware online client selection.

2. The online client selection method based on quality perception as described in claim 1, characterized in that, Prior to the step of calculating the data quality of the client in round t, the method further includes: Initialize client data and server; The client data includes settings for time slot length, client, number of channels, time slot loop variable, client loop variable, and channel loop variable.

3. The online client selection method based on quality perception as described in claim 1, characterized in that, The step of calculating the data quality of the client in round t includes: Calculate the heterogeneity of the client in round t based on the client's local training loss and testing loss; Calculate the attack level of the client in round t based on the client's global loss and test loss; The client data quality in round t is calculated by combining the heterogeneity and attack vulnerability of the client in round t.

4. The online client selection method based on quality perception as described in claim 2, characterized in that, The client data quality in round t The calculation formula is: in This indicates whether the client was selected to participate in the training process in round t-1. This represents the client data quality in round t-1. This represents the heterogeneity of the client in round t. This represents the attack level of the client in round t. This represents the heterogeneity weighting factor of the client in round t. It is the attack vulnerability weighting factor for the client in round t.

5. The online client selection method based on quality perception as described in claim 1, characterized in that, The communication latency reward valuation The calculation formula is: in, The average communication latency reward for the client to select a channel in round t-1. It is the first cumulative number of times the user selects a channel in round t-1; The estimated communication energy consumption reward The calculation formula is: in, The average communication energy consumption reward for the client to select a channel in round t-1. It is the first cumulative number of times the user selects a channel in round t-1.

6. The online client selection method based on quality awareness as described in claim 1, characterized in that, The training latency reward valuation The calculation formula is: in, Let be the average training latency reward for the client during round t-1. It represents the cumulative number of times the server selects a client in round t-1; The training energy consumption reward valuation The calculation formula is: in, Let be the average training energy reward for the client in round t-1. It represents the cumulative number of times the server selects a client in round t-1.

7. A quality-aware online client selection device, characterized in that, The device includes: The quality calculation module is used to calculate the data quality of the client in round t, where t is the time slot cycle variable; The communication valuation calculation module is used to calculate the communication latency reward valuation and the communication energy consumption reward valuation based on the cumulative number of times the user selects a channel in the (t-1)th round. The communication latency reward valuation and the communication energy consumption reward valuation are used to allocate a channel to each client. The training valuation module is used to calculate the training latency reward valuation and the training energy consumption reward valuation based on the cumulative number of times the server selects the client in round t-1. The client set calculation module is used to calculate the client set using the communication latency reward estimate, the communication energy consumption reward estimate, the training latency reward estimate, the training energy consumption reward estimate, and the data quality. The client set allocates channels to users and sets two vectors. and Both are the sum of the estimated communication latency reward and the estimated communication energy consumption reward; Let vector Sort in descending order and rank the channel performance among clients. , ; Let vector Sort clients in descending order and rank them by channel performance. , ; Channels are assigned to clients based on their performance ranking within the client, as defined in... The system prompts the client to select a channel for data transmission. If the value is 1, the client selects a channel for data transmission; otherwise, it does not. If multiple clients select the same channel, the channel is selected based on the client's performance ranking within the channel. If the operation is repeated, the next channel is selected for other clients based on the channel's performance ranking among the clients. The allocation module is used to recalculate channel allocations for the client set until all channels are allocated, in order to achieve quality-aware online client selection.

8. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing a computer program that is loaded and executed by the processor to implement the quality-aware online client selection method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which is loaded and executed by a processor to implement the quality-aware online client selection method as described in any one of claims 1-6.

Images