Federal learning dynamic client selection method based on dobby machine and fuzzy calculation

Abstract

The invention provides a federal learning dynamic client selection and privacy protection method based on a dobby machine and fuzzy calculation, and aims to improve the client selection efficiency and the privacy protection effect. According to the method, clients are dynamically selected through a UCB strategy in a dobby machine algorithm, rewards are calculated according to real-time performance (such as accuracy, training loss and the like) of each client, and updating is carried out in the training process. In combination with a fuzzy calculation technology, a Gaussian membership function is adopted to carry out fuzzy processing on uploaded model parameters, so that privacy is effectively protected, and the risk of data leakage is reduced. By introducing a dynamic adjustment factor, the method can balance the proportion of noise to original data, and further enhance privacy protection. Experimental results show that the client selection strategy based on the UCB shows relatively high training efficiency and model performance on MNIST and CIFAR-10 data sets, and is obviously superior to a traditional federated learning method. The method not only optimizes the client selection strategy, but also reduces the calculation and communication overhead, improves the practicability and adaptability of federal learning, and especially can effectively improve the training efficiency and privacy protection capability in a large-scale system.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and distributed computing, and specifically relates to a method for selecting and protecting a federated learning client based on a multi-armed bandit mechanism and fuzzy computing. Background Art

[0002] As an advanced distributed machine learning technology, federated learning aims to avoid the risk of centralized data storage and leakage by allowing multiple decentralized clients to perform training locally, greatly improving privacy protection. In federated learning, client selection is a key issue to improve training efficiency and system performance. The client selection strategy needs to consider multiple factors such as the client's computing power, communication bandwidth, and data quality to ensure that the global model can be trained efficiently. At present, the client selection method in federated learning mainly relies on classic optimization algorithms and model selection strategies. Relevant technical solutions include: Patent CN202410902829.4 proposes a dynamic client selection method based on a multi-armed bandit algorithm and Shapley value. This method optimizes client selection by calculating the contribution of each client; Patent CN202410879467.1 proposes a client selection method based on a non-steady-state multi-armed bandit and probabilistic Shapley value, which can dynamically adjust the client selection strategy to adapt to environmental changes; Patent CN202410283577.1 proposes a client scheduling method for hierarchical federated learning in a vehicle networking environment, which uses a confidence upper bound algorithm and data freshness to improve the fairness of scheduling; Patent CN202311369166.6 proposes a federated learning client selection method that combines privacy protection and incentive mechanism, and uses a combined multi-armed bandit model to evaluate the client learning quality.

[0003] However, existing technologies still have some problems in client selection: 1) Although a variety of algorithm combinations can provide optimized selection strategies for specific scenarios and applications, in actual applications, it is often necessary to repeatedly adjust parameters and models to obtain optimal performance. This process lacks general guidelines and therefore has high computational overhead; 2) Existing methods perform well in specific scenarios, but the algorithms lack versatility and adaptability, making them difficult to promote in different application scenarios, thus limiting the widespread application of federated learning technology.

[0004] Therefore, future research can be improved in the following aspects: first, improve the versatility of the client selection algorithm so that it can adapt to different network environments and application scenarios; second, explore more efficient dynamic adjustment strategies to reduce computing and communication overhead, thereby improving the practicality and efficiency of federated learning. In this context, the end-edge-cloud collaborative federated learning framework becomes an ideal solution. It can effectively reduce the system's computing burden and communication overhead by reasonably allocating computing and communication tasks between the cloud, edge, and client. In order to further optimize the client selection strategy, the Multi-Armed Bandit (MAB) algorithm is combined to provide an efficient dynamic decision-making mechanism, which can dynamically adjust the selection strategy according to the real-time performance of different clients. In addition, considering the needs of privacy protection and data security, fuzzy computing is used to fuzzify the model upload data, which can further improve privacy protection while maintaining model performance. This end-edge-cloud collaborative federated learning method based on multi-armed bandit and fuzzy computing can not only improve the flexibility and efficiency of client selection, but also reduce communication overhead and optimize the applicability of federated learning in different application scenarios. Summary of the invention

[0005] In view of the shortcomings of the existing technical methods, the purpose of the present invention is to provide a federated learning method that combines dynamic client selection with fuzzy computing. This method optimizes the client selection strategy, while enhancing the privacy protection effect through fuzzy computing, and improving the training efficiency and model performance of the federated learning system. To achieve the above purpose, the technical solution of the present invention is as follows:

[0006] A method for dynamic client selection of federated learning based on multi-armed bandit and fuzzy computing includes the following steps:

[0007] 1) First define the client set C = {C 1 ,C 2 ,...,C k}, where each client C i Participate in training based on its local data.

[0008] 2) Build a reward model: Each client calculates a reward based on its performance on the validation set (such as accuracy, training loss, etc.). This reward will be used for subsequent client selection and model updates.

[0009] 3) Use the UCB (Upper Confidence Bound) strategy to calculate the client's UCB value. The formula is: in: It is client C i The average reward in round t, n i (t) is the client Ci The number of times selected, t is the current round.

[0010] 4) According to the UCB value of each client, the client with the highest UCB value is selected to participate in the training.

[0011] 5) Training phase: The selected client is trained based on local data and calculates its performance on the local training set (such as loss function, accuracy, etc.). These results are fed back to the edge server as rewards.

[0012] 6) Fuzzy calculation processing upload process:

[0013] The model parameters uploaded by the client are fuzzy processed and the Gaussian membership function is used to fuzzify the data. The fuzzification formula is: i =β i ·μ(v;c i ,σ,α,θ i )+(1-β i )·μ(v;-c i ,σ,α,θ i ), where β i is the balance factor, used to adjust the influence of positive and negative centers, θ i It is the personalized periodic adjustment of each function. The fuzzified result is defuzzified to obtain the final uploaded data: Among them, w i is the weight, δ i is a nonlinear parameter that affects the contribution of the membership value. The fuzzified result is defuzzified to obtain the final uploaded data: Among them, r j is a nonlinear adjustment factor for each position. To enhance privacy protection, a dynamic adjustment factor γ is used to control the ratio of noise to original data: T ′ =T·e -γ +N·(1-e -γ ), where T is the original data, N is the noise, and γ controls the strength of the noise applied.

[0015] 7) Update the reward value based on the client's performance: For the UCB algorithm, update the client C i Reward value That is, the average reward: Among them, r i,j The jth selection of client C i The actual reward at the time.

[0016] 8) Dynamically adjust client selection:

[0017] The client selection strategy is adjusted by updating the reward value after each round of training. The system evaluates the performance of each client (such as training error, computing resource consumption, etc.) and dynamically updates the selection strategy to maintain training efficiency and privacy protection.

[0018] 9) Termination conditions and iterations:

[0019] The multi-armed bandit algorithm will continue to run until a preset number of training rounds or performance threshold is reached.

[0020] 10) Summary and model update:

[0021] All selected clients upload their updated model parameters to the edge server for aggregation, and multiple edge servers upload the model parameters to the cloud server for final aggregation, thereby optimizing the global model.

[0022] Compared with the prior art, the present invention has the following technical features and effects:

[0023] (1) The present invention combines the multi-armed bandit algorithm (MAB) and fuzzy computing to propose a dynamic client selection strategy that can dynamically adjust the selection strategy according to the real-time performance of each client. Different from the traditional method that relies on fixed algorithms and parameter adjustments, the present invention dynamically updates the client selection by introducing the UCB strategy, thereby improving the flexibility and accuracy of client selection and reducing manual intervention and computing overhead.

[0024] (2) The present invention uses fuzzy computing technology to fuzzify the model parameters uploaded by the client, and uses Gaussian membership function to protect the privacy of the model data. This process can effectively reduce the risk of data leakage and maintain the performance and accuracy of the model. Compared with the existing technology that only protects privacy through encryption, the present invention provides additional privacy protection for data through fuzzification technology, further improving the privacy protection effect.

[0025] (3): The client selection and fuzzy calculation method of the present invention can effectively reduce the computation and communication overhead, especially in large-scale federated learning systems. By optimizing the efficiency of client selection and reducing unnecessary data transmission, the present invention can significantly reduce the resource consumption of the system, improve the overall training efficiency, and adapt to different application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 The structural diagram used in the present invention

[0027] Figure 2 Workflow diagram used in the present invention

[0028] Figure 3 The training results of the present invention on the MNIST data set are

[0029] Figure 4 This is the training result of the present invention on the CIFAR-10 dataset DETAILED DESCRIPTION

[0030] The present invention is described below in conjunction with the accompanying drawings and embodiments.

[0031] The present invention proposes a method for selecting a dynamic client for federated learning based on a combination of a multi-armed bandit and fuzzy computing. The specific structure is as follows: Figure 1 This method aims to dynamically select clients participating in federated learning by integrating fuzzy computing and multi-armed bandit algorithms, thereby improving the privacy protection, training efficiency and robustness of the global model. The technical solution is as follows:

[0032] Step 1: Initialization phase

[0033] The cloud server initially selects all clients and sends the current global model to all edge servers, which then send the global model to all clients. After receiving the global model, the client trains the model based on local data and uploads it for update.

[0034] Step 2: Client model update

[0035] Define client set C = {C 1 ,C 2 ,...,C k}, where each client C i Participate in training based on its local data. Each client calculates a reward based on its performance on the validation set (such as accuracy, training loss, etc.). This reward will be used for subsequent client selection and model updates.

[0036] Step 3: Client selection

[0037] Use the UCB (Upper Confidence Bound) strategy to calculate the client's UCB value. The formula is: in: It is client C i The average reward in round t, n i (t) is the client C i The number of times selected, t is the current round. According to the UCB value of each client, the client with the highest UCB value is selected to participate in the training.

[0038] Step 4: Training phase

[0039] The selected client is trained based on local data and calculates its performance on the local training set (such as loss function, accuracy, etc.). These results are fed back to the edge server as rewards.

[0040] Step 5: Fuzzy calculation processing upload process

[0041] The model parameters uploaded by the client are fuzzy processed and the Gaussian membership function is used to fuzzify the data. The fuzzification formula is:

[0042] m i =β i ·μ(v;c i ,σ,α,θ i )+(1-β i )·μ(v;-c i ,σ,α,θ i ),

[0043] Among them, β i is the balance factor, used to adjust the influence of positive and negative centers, θ i It is the personalized periodic adjustment of each function.

[0044] Defuzzify the fuzzified result to get the final uploaded data:

[0045]

[0046] Among them, w i is the weight, δ i is a nonlinear parameter that affects the contribution of the membership value.

[0047] Defuzzify the fuzzified result to get the final uploaded data:

[0048]

[0049] Among them, r j is a nonlinear adjustment factor for each position.

[0050] To enhance privacy protection, a dynamic adjustment factor γ is used to control the ratio of noise to original data:

[0051] T ′ =T·e -γ +N·(1-e -γ ),

[0052] Where T is the original data, N is the noise, and γ controls the strength of the noise applied.

[0053] Step 6: Update the reward value

[0054] For the UCB algorithm, update client C i Reward value That is, the average reward:

[0055]

[0056] Among them, r i,j The jth selection of client C i The actual reward at the time.

[0057] Step 7: Dynamically adjust client selection

[0058] The client selection strategy is adjusted by updating the reward value after each round of training. The system evaluates the performance of each client (such as training error, computing resource consumption, etc.) and dynamically updates the selection strategy to maintain training efficiency and privacy protection.

[0059] Step 8: Termination Conditions and Iterations

[0060] The multi-armed bandit algorithm will continue to run until a preset number of training rounds or performance threshold is reached.

[0061] Step 9: Summary and model update

[0062] All selected clients upload their updated model parameters to the edge server for aggregation, and multiple edge servers upload the model parameters to the cloud server for final aggregation, thereby optimizing the global model.

[0063] The algorithm structure diagram proposed by the present invention is as follows: Figure 4 As shown in the figure, this system consists of 50 clients, 5 edge servers and 1 cloud server. Each edge server manages the same number of clients and ensures the consistency of the data volume of each client during the training process. The MNIST and CIFAR-10 datasets are used as tests. The experiment compares the performance of three different client selection mechanisms in the edge-cloud federated learning architecture: basic federated learning (HF), client selection asynchronous hierarchical federated averaging (CSA-HF), and client selection based on multi-armed bandit UCB (MAB-UCB-HF).

[0064] After analyzing the experimental data, Figure 3 The training results of the MNIST dataset after aggregation in the cloud are shown. Figure 4The training results of the CIFAR-10 dataset after aggregation in the cloud are shown. When evaluating the MNIST dataset, all models quickly reach high accuracy within the first 20 training cycles, with an accuracy of or above 0.9. In the middle and late stages of training, especially between 80 and 100 rounds, the MAB-UCB-HF ​​model performs slightly better, maintaining an average accuracy of about 0.97. In contrast, the accuracy of the HF model stabilizes at around 0.96, while the accuracy of the CSA-HF model fluctuates between 0.95 and 0.97. It is worth noting that in the last 10 training cycles, the MAB-UCB-HF ​​model showed higher stability, with an accuracy that remained above 0.97, while the other two models experienced larger fluctuations.

[0065] For the more complex CIFAR-10 dataset, the MAB-UCB-HF ​​model also showed remarkable stability and superiority throughout the training process. In the early stages of training (rounds 0 to 50), the accuracy of the MAB-UCB-HF ​​and CSA-HF models rose almost synchronously, with MAB-UCB-HF ​​slightly ahead. Entering the late stages of training (rounds 250 to 300), the accuracy of MAB-UCB-HF ​​remained between 0.75 and 0.77, while the accuracy of the HF and CSA-HF models fluctuated between 0.72 and 0.76. In particular, in the zoomed-in view of rounds 240 to 300, the MAB-UCB-HF ​​model demonstrated its ability to continue learning on complex datasets.

[0066] Through detailed analysis of the data points, it is clear that the MAB-UCB-HF ​​model outperforms the HF model and the CSA-HF model on both the MNIST and CIFAR-10 datasets, especially in terms of stability and persistence of test accuracy. This excellent performance is due to its effective client selection strategy based on the upper confidence bound (UCB) multi-armed bandit strategy. This strategy optimizes client participation, improves learning efficiency and result quality, and maintains high performance even when the number of available clients is limited. These findings highlight the broad potential and value of this method in practical applications, especially in scenarios that require high precision and complex data processing.

[0067] The advantages of this invention are that it dynamically adjusts client selection through a multi-armed bandit algorithm, reduces manual intervention and computing overhead, and improves flexibility and efficiency; at the same time, it uses fuzzy computing to enhance privacy protection and more effectively reduce the risk of data leakage than traditional encryption methods. In addition, this method can effectively reduce the computing and communication overhead in large-scale federated learning systems and improve overall training efficiency.

[0068] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention. For example, the specific selection of technical details, implementation methods or algorithms, etc., may be adjusted or optimized according to actual application requirements. All of these modifications or equivalent replacements should be included in the scope of the claims of the present invention.

Claims

1. A method for dynamic client selection of federated learning based on multi-armed bandit and fuzzy computing, comprising the following steps: 1) First define the client set C = {C1, C2, ..., C k }, where each client C i Participate in training based on its local data. 2) Build a reward model: Each client calculates a reward based on its performance on the validation set (such as accuracy, training loss, etc.). This reward will be used for subsequent client selection and model updates. 3) Use the UCB (Upper Confidence Bound) strategy to calculate the client's UCB value. The formula is: in: It is client C i The average reward in round t, n i (t) is the client C i The number of times selected, t is the current round. 4) According to the UCB value of each client, the client with the highest UCB value is selected to participate in the training. 5) Training phase: The selected client is trained based on local data and calculates its performance on the local training set (such as loss function, accuracy, etc.). These results are fed back to the edge server as rewards. 6) Fuzzy calculation processing upload process: The model parameters uploaded by the client are fuzzy processed and the Gaussian membership function is used to fuzzify the data. The fuzzification formula is: m i =b i ·μ(v;c i , s, a, θ i )+(1-β i )·μ(v;-c i , s, a, θ i ), Among them, β i is the balance factor, used to adjust the influence of positive and negative centers, θ i It is the personalized periodic adjustment of each function. The fuzzified result is defuzzified to obtain the final uploaded data: Among them, w i is the weight, δ i is a nonlinear parameter that affects the contribution of the membership value. Defuzzify the fuzzified result to get the final uploaded data: Among them, r j is a nonlinear adjustment factor for each position. To enhance privacy protection, a dynamic adjustment factor γ is used to control the ratio of noise to original data: T′=T·e -γ +N·(1-e -γ ), Where T is the original data, N is the noise, and γ controls the strength of the noise applied. 7) Update the reward value based on the client's performance: For the UCB algorithm, update the client C i Reward value That is, the average reward: 8) Among them, r i，j The jth selection of client C i The actual reward at the time. 8) Dynamically adjust client selection: The client selection strategy is adjusted by updating the reward value after each round of training. The system evaluates the performance of each client (such as training error, computing resource consumption, etc.) and dynamically updates the selection strategy to maintain training efficiency and privacy protection. 9) Termination conditions and iterations: The multi-armed bandit algorithm will continue to run until a preset number of training rounds or performance threshold is reached. 10) Summary and model update: All selected clients upload their updated model parameters to the edge server for aggregation, and multiple edge servers upload the model parameters to the cloud server for final aggregation, thereby optimizing the global model.

Images