An edge-computing-oriented federated learning client dynamic screening and joint incentive optimization system

Abstract

The application discloses a kind of federated learning client dynamic screening and joint incentive optimization systems for edge computing, comprising: dynamic quality perception screening module, two-way incentive collaboration module and light weight strategy solving module;Dynamic quality perception screening module is used to intelligently screen target value client participation training according to the data quality of each client before federated iteration;Two-way incentive collaboration module is used to solve the optimal strategy combination using light weight strategy solving module based on aggregator and screened client, and the aggregator and client decision are optimized according to the optimal strategy combination;Light weight strategy solving module is used to calculate the optimal reward strategy of aggregator and the optimal data contribution strategy of each client, and obtain the optimal strategy combination.The application establishes the incentive interaction between aggregator and client represented by edge device in edge scene, improves global model accuracy in the case of unknown client data quality, and avoids incentive unfairness.

Description

Technical Field

[0001] This invention belongs to the fields of edge computing and federated learning technology, and particularly relates to a dynamic screening and joint incentive optimization system for federated learning clients for edge computing. Background Technology

[0002] Edge computing, by deploying computing, storage, and network capabilities close to the data source at the network edge, aims to meet the demands for low latency, high bandwidth, localized data processing, and security and privacy. It is widely used in real-time status monitoring of factory production line equipment, real-time traffic flow analysis of roadside units, and personalized services for wearable devices. In recent years, with the deep integration of the Internet of Things (IoT), 5G / 6G, and artificial intelligence (AI), massive amounts of data are generated in these scenarios. Utilizing this data for machine learning model training is core to achieving intelligent decision-making. This places demands on ultra-low latency, high bandwidth, and high reliability for data processing and intelligent decision-making. However, traditional centralized cloud training models struggle to meet these requirements due to long data transmission distances and network congestion. Specifically, regarding data privacy protection, edge data often contains sensitive information, such as core factory process parameters and personal biometrics. With increasingly stringent data privacy regulations, traditional centralized cloud training carries the risk of raw data leakage due to its large-scale data transmission. In terms of network bandwidth, continuously uploading petabyte / electronabyte (PB / EB) level raw edge data, especially video, images, and high-frequency sensor streams, to the cloud consumes enormous bandwidth resources, is costly, and is impractical or extremely inefficient in remote factories and scenarios with limited mobile cellular network bandwidth. Regarding the utilization of edge device resources, modern edge devices, such as embedded systems, industrial controllers, smartphones, and roadside units, while typically severely limited in computing power, memory, storage space, and battery life, possess a certain level of computing capability (CPU / GPU / NPU). Centralized models have failed to effectively utilize this distributed computing potential. These challenges severely restrict the real-time performance, security, and cost-effectiveness of edge intelligence applications, urgently requiring new distributed machine learning paradigms to overcome these bottlenecks.

[0003] Federated learning, as a distributed machine learning paradigm, provides an effective way to address the aforementioned pain points of edge computing. Its core innovation lies in the fact that the original training data is retained on the local edge device. The training process is achieved through iterative steps:

[0004] Model distribution: Aggregators deployed in edge clouds or regional data centers distribute the current global model to selected edge clients.

[0005] Local training: The client uses local private data and local computing resources to train on the basis of the global model and calculate update parameters such as model gradient and model weight difference.

[0006] Secure upload: The client only encrypts and uploads the model update results (not the original data) to the central coordination node.

[0007] Secure aggregation: The aggregator securely aggregates multiple received model update results to generate a new global model.

[0008] Model update and iteration: A new global model is distributed, and the next round of iteration begins.

[0009] Federated learning's technological advantages overcome the difficulties of applying traditional centralized learning in edge computing scenarios. In predictive maintenance within the Industrial Internet of Things (IIoT), federated learning enables each device to train its fault prediction model locally, sharing only model updates. This protects core factory data, reduces data transmission across production lines / plants, and leverages the device's own computing power. In vehicle-to-everything (V2X) collaborative perception and decision-making, federated learning allows roadside units to train their perception models locally or contribute local learning results, enabling rapid environmental awareness updates and collaborative decision-making, avoiding bandwidth explosions and latency caused by uploading raw video / point cloud data. In smart city real-time monitoring and response, federated learning allows cameras or edge servers to analyze video streams locally, training or updating object detection or behavior recognition models, sharing only key model updates or event summaries, protecting privacy and improving response speed.

[0010] In federated learning, the accuracy of the global model is typically related to the total amount and quality of data in the client datasets participating in the federated learning process. Data quality refers to the difference between the distribution of individual labels in the client's data and the total label distribution of all participating clients' data. A greater difference indicates a higher degree of non-independent and identically distributed (Io-distributed) nature in the client datasets, resulting in lower data quality. This is because local models trained on datasets with varying degrees of Io-distribution typically reduce the accuracy of the global model during aggregation. In common edge computing scenarios, due to differences in data collection time, space, and sensors, the degree of Io-distribution varies significantly among client datasets. Therefore, allowing clients to participate in federated learning without selection often fails to improve the accuracy of the global model and may even reduce it. More importantly, the participation of low-quality clients wastes the limited power, computation time, and network bandwidth of edge devices, exacerbating the ineffective consumption of edge resources. However, traditional federated learning frameworks typically assume that the quality of client data is known to the aggregator. This is difficult to achieve in practical applications of federated learning because of its client-side data privacy protection nature. The aggregator cannot directly observe the client's original dataset and therefore cannot assess the client's data quality, thus failing to select clients for participation based on that quality. Furthermore, client-side training of local models incurs costs such as communication and computational energy. If the aggregator does not provide sufficient compensation, clients will not actively contribute data to federated learning, leading to a decrease in the total amount of data used for federated learning and consequently reducing the accuracy of the global model.

[0011] In summary, the design of incentive mechanisms for federated learning with client selection still faces three challenges: First, data quality is unknown. In practical edge-side federated learning applications, due to the privacy protection mechanism of federated learning, the degree of non-independent and identically distributed nature of client datasets cannot be directly evaluated. This means that data quality is usually unknown to the aggregator, leading to poor filtering performance, wasted resources, and increased accuracy loss. Second, global model accuracy evaluation is distorted. Incentive mechanisms designed to maximize global model accuracy typically require a global model accuracy evaluation function. However, due to the non-convex nature of machine learning model loss functions, pre-defined accuracy proxy functions (such as logarithmic loss) cannot reflect the true improvement, resulting in inaccurate incentive allocation and insufficient rewards for high-quality updates. Third, policy solving is infeasible. When calculating the optimal policies for the aggregator and clients under information asymmetry using heuristic search or reinforcement learning, the optimality of the obtained policies cannot be guaranteed, leading to global model accuracy loss. Furthermore, the edge-side aggregator lacks sufficient computing power to generate reliable policies in real time, resulting in a high risk of system failure. Summary of the Invention

[0012] To address the aforementioned technical issues, this invention proposes a dynamic screening and joint incentive optimization system for federated learning clients in edge computing. It establishes an incentive interaction between the aggregator and clients represented by edge devices in edge scenarios, improving the accuracy of the global model even when the data quality of the clients is unknown, and avoiding incentive unfairness.

[0013] To achieve the above objectives, this invention provides a dynamic screening and joint incentive optimization system for federated learning clients for edge computing, comprising: a dynamic quality-aware screening module, a bidirectional incentive collaboration module, and a lightweight strategy solving module;

[0014] The dynamic quality-aware filtering module is used to intelligently filter target value clients to participate in training based on the data quality of each client before the federated iteration.

[0015] The bidirectional incentive collaboration module is used to solve for the optimal strategy combination based on the aggregator and the selected clients using the lightweight strategy solving module, and to optimize the decisions of the aggregator and clients based on the optimal strategy combination.

[0016] The lightweight strategy solving module is used to calculate the optimal reward strategy of the aggregator and the optimal data contribution strategy of each client, and obtain the optimal strategy combination.

[0017] Optionally, the dynamic quality-aware screening module includes: a lightweight gradient difference analyzer and an online quality tracker;

[0018] The lightweight gradient difference analyzer is used to select the K clients with the highest data quality to participate in federated learning based on the historical quality information of each client.

[0019] The online quality tracker is used to record the data quality information of the clients participating in this round of training, update the screening criteria for each client, and use it for the next round of client screening.

[0020] Optionally, the selection of the K clients with the highest data quality to participate in federated learning based on the historical quality information of each client includes:

[0021] The historical quality information of the clients is sorted in descending order, and the K clients with the highest data quality are selected from the sorted clients to participate in federated learning.

[0022] Optionally, data quality information from clients participating in this training round can be recorded, and the selection criteria for each client can be updated for use in the next round of client selection, including:

[0023] Update the cumulative number of times each client has participated in federated learning up to the t-th round of global iteration;

[0024] Based on the cumulative number of times participating in federated learning, update the average data quality of each client up to the t-th round of global iteration;

[0025] Based on the average data quality, update the CUCB-based data quality of each client in round t;

[0026] Based on the quality of the CUCB-based data, the screening criteria for each client are updated.

[0027] Optionally, optimizing the aggregator and client decisions based on the optimal strategy combination includes:

[0028] Step 1: Set an upper limit on the number of game iterations (iter) max Randomly initialize the aggregator reward decision;

[0029] Step 2: Each client calculates the optimal data contribution strategy within the current game iteration using the lightweight strategy solving module, based on the reward strategy initialized by the aggregator.

[0030] Step 3: The aggregator calculates the optimal reward strategy within the current game iteration based on the optimal data contribution strategy of each client within the current game iteration through the lightweight strategy solving module.

[0031] Step 4: Repeat steps 2 and 3 until the maximum number of iterations is reached. max The iteration continues until the strategies of the aggregator and each client obtained in the current iteration are consistent with the strategies obtained in the previous iteration, at which point the iteration stops and the optimized aggregator and client decisions are obtained.

[0032] Optionally, each client calculates the optimal data contribution strategy using the lightweight strategy solving module based on the reward strategy initialized by the aggregator, including:

[0033] Step 1: Determine if the reward strategy after aggregator initialization is 0. If it is 0, the optimal data contribution strategy for each client is 0; otherwise, proceed to Step 2.

[0034] Step 2: Determine whether the reward strategy after aggregator initialization is less than the judgment value. If it is less than, proceed to step three; otherwise, determine the optimal data contribution strategy for the i-th client.

[0035] Step 3: The reward strategy after aggregator initialization is less than the judgment value. For clients, the optimal data contribution strategy is in, For the optimal contribution strategy, τ t The reward strategy after the aggregator is initialized. Let λ represent the actual data quality of the i-th client in this round of global iteration, and λ be the worst data quality that the aggregator can tolerate. i Let D be the cost coefficient for the i-th client. i Let be the total amount of data for the i-th client.

[0036] Optionally, the aggregator calculates the optimal reward strategy within the current game iteration based on the optimal data contribution strategy of each client using the lightweight strategy solving module, including:

[0037] Step 1: Calculate the parameters for each client.

[0038] Step 2: Obtain the maximum amount of data that each client can contribute, D1, D2, ..., D. K , will the client according to The values ​​are sorted in descending order;

[0039] Step 3: Based on the clients sorted in descending order, divide the range of values ​​for the aggregator reward decision;

[0040] Step 4: Calculate the locally optimal reward for each range of values ​​in the aggregator reward strategy;

[0041] Step 5: Based on the local optimal reward, obtain the global optimal reward strategy within the current game iteration.

[0042] Optionally, the range of values ​​for the aggregator reward decision can be divided as follows:

[0043]

[0044] when At that time, the locally optimal reward is:

[0045]

[0046] when At that time, local optimal reward

[0047] when At that time, local optimal reward in, For locally optimal reward, To be according to After sorting in descending order, the first one is ranked D in i , To be according to After sorting in descending order, the last one is ranked D in i , To be according to After sorting in descending order, the first one is ranked In To be according to After sorting in descending order, the last one is ranked In ν1 and v2 are balancing parameters in the aggregator utility function used to balance the aggregator's emphasis on global model accuracy and reward cost, φ tH For precision control parameters, For precision control parameters, w i For the i-th client, For a client collection, w j For the j-th client, ψ is the precision control parameter. Here, H represents the intermediate calculation parameters, and H represents the number of local iterations during federated learning training on each client. This is the set of clients selected in the t-th round of global iteration.

[0048] Compared with the prior art, the present invention has the following advantages and technical effects:

[0049] This invention achieves the following breakthrough effects in edge federated learning through a joint optimization mechanism that combines dynamic quality-aware screening with bidirectional incentive collaboration:

[0050] (1) Significantly improved global model accuracy: High-quality clients are dynamically selected through gradient difference analysis, reducing the participation rate of low-quality clients and improving model accuracy under the same number of iterations. The incentive module incorporates data quality and resource cost into the utility function, driving clients to contribute high-quality data, thereby improving the model convergence speed per unit amount of data.

[0051] (2) Significantly optimized edge resource consumption: The CUCB mechanism of the quality-aware module reduces invalid exploration, saves client computing power, and extends battery device life. Lightweight solution strategy reduces aggregator CPU utilization and adapts to resource-constrained edge nodes.

[0052] (3) Privacy Protection and Fairness Guarantee: Data quality is inferred from model gradient differences throughout the process (without requiring access to the original data), thus meeting privacy protection requirements. The incentive mechanism addresses the fairness issue of "high-quality data with low compensation" in traditional solutions by implementing quality-based payment.

[0053] (4) Enhanced system robustness: The uniqueness proof of the equilibrium in the Steinberg game guarantees the stability of policy convergence, reducing policy oscillations compared to traditional reinforcement learning methods. It supports dynamic edge environments (random device offline, network fluctuations), reducing the training task interruption rate. Attached Figure Description

[0054] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0055] Figure 1 This is a structural diagram of a federated learning client dynamic screening and joint incentive optimization system for edge computing according to an embodiment of the present invention. Detailed Implementation

[0056] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0057] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0058] To address the shortcomings of traditional federated learning paradigms in edge computing scenarios, this invention constructs a mapping relationship between the data distribution of each client dataset and the global model convergence performance, based on real-time local model updates from clients and analysis of global model convergence. This is done even when the specific details of the client datasets are unknown. Furthermore, an incentive mechanism is designed to encourage active client participation in federated learning, aligning the behavior of both the aggregator and clients towards global model convergence. This maximizes global model accuracy by incentivizing client participation even with unknown client data quality. This invention selects clients with high-quality data for federated learning and incentivizes them to contribute more data, thus overcoming the challenge of optimizing global model accuracy when client data quality is unknown.

[0059] This embodiment proposes a dynamic selection and joint incentive optimization system for federated learning clients in edge computing, such as... Figure 1 As shown, it specifically includes: a dynamic quality perception screening module, a two-way incentive collaboration module, and a lightweight strategy solving module;

[0060] The dynamic quality-aware filtering module is used to intelligently filter target value clients to participate in training based on the data quality of each client before federated iteration;

[0061] The bidirectional incentive collaboration module is used to solve for the optimal strategy combination based on the aggregator and the selected clients using the lightweight strategy solving module, and optimize the decisions of the aggregator and clients based on the optimal strategy combination.

[0062] The lightweight strategy solving module is used to calculate the optimal reward strategy for the aggregator and the optimal data contribution strategy for each client, and obtain the optimal strategy combination.

[0063] Specifically, the dynamic quality-aware screening module: Before each round of federated iterations, this module intelligently selects high-value clients to participate in training based on the data quality of each client. First, a lightweight gradient dissimilarity analyzer is built to dynamically generate data quality assessment values ​​based on the difference between the client's historical local gradient model and the global gradient model. Next, an online quality tracker is designed to record the actual quality performance of each client after participation, creating a client quality history archive. Through these methods, privacy-preserving evaluation is achieved, requiring no access to the original data throughout the process, relying only on model gradient updates.

[0064] (2) Two-way Incentive Collaboration Module: To ensure that the actions of both the aggregator and the client contribute to improving the accuracy of the global model, this module, based on Steinberg game theory and federated learning convergence analysis, synchronously optimizes the decisions of the aggregator and the client through a reward mechanism. This module adopts a hierarchical decision-making framework, with the aggregator acting as the leader, controlling the reward budget to maximize global accuracy; and the client acting as the follower, adjusting its data contribution based on the reward. A joint accuracy-cost utility function is designed, where the aggregator's utility is the global model accuracy evaluation function obtained based on federated learning convergence analysis minus the reward expenditure, and the client's utility is the reward income minus the local resource consumption cost. By establishing a game-driven convergence mechanism, system equilibrium is achieved through two-way policy adjustment.

[0065] (3) Lightweight Strategy Solving Module: The goal of this module is to efficiently compute the optimal strategy combination. It consists of a client-side strategy parser and an aggregator strategy optimizer. The client-side strategy parser directly calculates the optimal data contribution of each client using gradient backpropagation, while the aggregator strategy optimizer uses an efficient numerical iterative algorithm to solve for the reward allocation scheme. The computational complexity of the solving algorithm is controlled to O(n), adapting to the computing power of the edge aggregator.

[0066] Furthermore, the dynamic quality-aware screening module includes: a lightweight gradient difference analyzer and an online quality tracker;

[0067] A lightweight gradient difference analyzer is used to select the K clients with the highest data quality to participate in federated learning based on the historical quality information of each client.

[0068] An online quality tracker is used to record the data quality information of the clients participating in this round of training, update the screening criteria for each client, and use it for the next round of client screening.

[0069] Furthermore, the process of selecting the K clients with the highest data quality to participate in federated learning based on the historical quality information of each client includes:

[0070] The historical quality information of the clients is sorted in descending order, and the K clients with the highest data quality are selected from the sorted clients to participate in federated learning.

[0071] Furthermore, the data quality information of the clients participating in this round of training is recorded, and the selection criteria for each client are updated for the next round of client selection, including:

[0072] Update the cumulative number of times each client has participated in federated learning up to the t-th round of global iteration;

[0073] Based on the cumulative number of times participating in federated learning, update the average data quality of each client up to the t-th round of global iteration;

[0074] Based on the average data quality, update the UCB-based data quality of each client in round t;

[0075] Based on UCB-based data quality, update the screening criteria for each client.

[0076] Furthermore, optimizing aggregator and client decisions based on the optimal strategy combination includes:

[0077] Step 1: Set an upper limit on the number of game iterations (iter) max Randomly initialize the aggregator reward decision;

[0078] Step 2: Each client calculates the optimal data contribution strategy within the current game iteration using the lightweight strategy solving module, based on the reward strategy initialized by the aggregator.

[0079] Step 3: The aggregator calculates the optimal reward strategy within the current game iteration based on the optimal data contribution strategy of each client within the current game iteration through the lightweight strategy solving module.

[0080] Step 4: Repeat steps 2 and 3 until the maximum number of iterations is reached. max The iteration continues until the strategies of the aggregator and each client obtained in the current iteration are consistent with the strategies obtained in the previous iteration, at which point the iteration stops and the optimized aggregator and client decisions are obtained.

[0081] Furthermore, each client calculates the optimal data contribution strategy based on the reward strategy initialized by the aggregator, using the lightweight strategy solving module, including:

[0082] Step 1: Determine if the reward strategy after aggregator initialization is 0. If it is 0, the optimal data contribution strategy for each client is 0; otherwise, proceed to Step 2.

[0083] Step 2: Determine whether the reward strategy after aggregator initialization is less than the judgment value. If it is less than, proceed to step three; otherwise, determine the optimal data contribution strategy for the i-th client.

[0084] Step 3: The reward strategy after aggregator initialization is less than... For clients, the optimal data contribution strategy is in, For the optimal contribution strategy, τ t The reward strategy after the aggregator is initialized. Let λ represent the actual data quality of the i-th client in this round of global iteration, and λ be the worst data quality that the aggregator can tolerate. i Let D be the cost coefficient for the i-th client. i Let be the total amount of data for the i-th client.

[0085] Furthermore, based on the optimal data contribution strategy of each client within the current game iteration, the aggregator calculates the optimal reward strategy within the current game iteration using a lightweight strategy solving module, including:

[0086] Step 1: Calculate the parameters for each client.

[0087] Step 2: Obtain the maximum amount of data that each client can contribute, D1, D2, ..., D. K , will the client according to The values ​​are sorted in descending order;

[0088] Step 3: Based on the clients sorted in descending order, divide the range of values ​​for the aggregator reward decision;

[0089] Step 4: Calculate the locally optimal reward for each range of values ​​in the aggregator reward strategy;

[0090] Step 5: Based on the local optimal reward, obtain the globally optimal reward strategy within the current game iteration.

[0091] This embodiment will be illustrated using federated learning for quality prediction in the automotive manufacturing supply chain as an example:

[0092] A vehicle manufacturer wants to build a high-precision component defect prediction model to identify high-risk batches before they enter the assembly line, reducing rework costs. This model requires real-time production data from multiple key component suppliers (e.g., Company A provides engine blocks, Company B provides electronic control systems, and Company C provides braking components) for training. However, training this model faces multiple challenges: data sovereignty and privacy—each supplier's production data (such as sensor readings, process parameters, and quality inspection records) is their core competitiveness and trade secret, and suppliers strongly oppose directly sharing raw data with the vehicle manufacturer or other suppliers; data quality heterogeneity—each component supplier collects data in different ways and under different scenarios, resulting in inconsistent data quality; incentive requirements—participating in federated learning consumes the supplier's own resources (computing power, electricity costs, network bandwidth, engineer maintenance time), requiring sufficient economic compensation or commercial benefits from the vehicle manufacturer; otherwise, there is no incentive to contribute data or train diligently; "free-riding" risk—without reasonable incentives, high-quality supplier A may reduce its contribution (due to high costs), while low-quality supplier B may try to contribute more low-value data to fraudulently obtain rewards.

[0093] In this embodiment, a global defect prediction model aggregator is deployed on the edge cloud of the vehicle manufacturer. Local factory edge nodes of each supplier act as clients, storing private data and training local defect prediction models. To collaboratively train a global component defect prediction model, in each global iteration: the aggregator first selects the K clients with the highest data quality to participate in federated learning, based on historical client data quality information and using a lightweight gradient difference analyzer in the dynamic quality-aware screening module. The selected clients and the aggregator jointly enter a bidirectional incentive collaboration module, forming a Steinberg game. A lightweight strategy solving module iteratively calculates the aggregator's optimal reward calculation strategy and each client's optimal data contribution strategy until a game equilibrium is reached. Next, the aggregator distributes the parameters of the current global defect prediction model to the selected clients. Each client trains a local defect prediction model based on its own optimal data contribution strategy and the global model parameters. After training, the local model parameters are uploaded back to the aggregator. After completing the global model aggregation, the aggregator pays rewards to each participating client. Finally, the online quality tracker in the dynamic quality-aware screening module records the actual data quality of the clients participating in federated learning in this round and updates the quality parameters of each client for the next round of global iteration.

[0094] (1) Dynamic Quality-Aware Screening Module: This module consists of a lightweight gradient difference analyzer and an online quality tracker. When screening clients to participate in federated learning, the lightweight gradient difference analyzer selects the K clients with the highest data quality based on the historical data quality information of each client's previous participation in federated learning. After the selected clients complete this round of federated learning, the online quality tracker records the data quality information of these clients in this round and updates the screening criteria for each client for the next round of client screening.

[0095] Before the start of the t-th round of federated learning global iteration (assuming there are a total of T rounds of global iteration), the specific execution process of the lightweight gradient difference analyzer is as follows:

[0096] Step 1: Assign each client data according to its respective CUCB-based data quality. (Assuming there are N clients) sorted in descending order, where CUCB-based data quality is based on the upper confidence bound of the data quality. The CUCB algorithm is a combined upper confidence bound algorithm, and this CUCB-based data quality is a key intermediate variable in the process of executing this algorithm.

[0097] Step 2: Select the top K clients from the clients sorted in descending order to participate in federated learning.

[0098] After the t-th round of global federated learning iterations, the specific execution process of the online quality tracker is as follows:

[0099] Step 1: Update the cumulative number of times each client has participated in federated learning up to the t-th round of global iteration. The updated formula is as follows:

[0100]

[0101] in, This is an indicator function that outputs 1 when the function input is true, and 0 otherwise.

[0102] Step 2: Update the average data quality of each client up to the t-th round of global iteration. The updated formula is as follows:

[0103]

[0104] in, This represents the actual data quality of the i-th client in this round of global iteration. In fact, at the end of the t-th round of global iteration, the aggregator can only derive the actual data quality of the clients participating in federated learning using the following formula:

[0105]

[0106] in, Let g represent the gradients of the global model and the i-th client-side local model, respectively. max This represents the maximum expected label loss. This represents the bulldozer distance of the i-th client dataset.

[0107] Step 3: Update the UCB-based data quality for each client in round t. The updated formula is:

[0108]

[0109] CUCB-based data quality serves as the basis for the lightweight gradient difference analyzer to select clients. This data quality ensures that clients that are selected less often have a greater chance of being selected in future global iterations, thus avoiding missing out on clients with high data quality, which is optimism in the face of uncertainty.

[0110] (2) Two-way incentive collaboration module: After selecting clients to participate in federated learning in the t-th round of global iteration, this module uses the aggregator as the leader and the selected clients as followers, forming a two-stage Steinberg game incentive mechanism. After the aggregator gives its reward strategy, the selected clients determine their data contribution strategy according to the reward strategy, and this process is repeated until the game equilibrium is reached.

[0111] Game equilibrium is defined as a situation where neither the leader nor the followers can improve their own gains by changing only their own decisions without altering the strategies of other participants. Specifically, in a federated learning system, the aggregator cannot improve its utility (global model accuracy minus reward cost) by simply changing its reward decisions, because if the aggregator changes its reward decisions, the clients, as followers, will also change their data contribution strategies. Similarly, clients cannot improve their utility (reward gain minus resource cost) by simply changing their reward decisions.

[0112] The utility function of the tq-th aggregator is as follows:

[0113] Taking a federated learning system with one aggregator and K selected clients as an example, the utility function of the aggregator is:

[0114]

[0115] The first term is the global model accuracy evaluation function derived from federated learning convergence analysis, and the second term is the reward paid by the aggregator to each client. The aggregator adjusts its reward strategy τ. t To influence the data contribution strategies of each client. This maximizes the value of its utility function.

[0116] The utility function for the i-th client is:

[0117]

[0118] The first item is the client's reward, and the second is its resource consumption cost. The client adjusts its data contribution strategy accordingly. To influence the aggregator's reward strategy τ t This maximizes the value of its utility function.

[0119] The specific two-way incentive and collaboration method is as follows:

[0120] Step 1: Set an upper limit on the number of game iterations (iter) max Randomly initialize the aggregator reward decision τ t .

[0121] Step 2: Each client calculates the optimal data contribution strategy using the lightweight strategy solving module, based on the aggregator's reward strategy.

[0122] Step 3: The aggregator calculates the optimal reward strategy based on the data contribution strategy of each client through the lightweight strategy solving module.

[0123] Step 4: Repeat steps 2 and 3 until the maximum number of iterations is reached. max This continues until the strategies of the aggregator and each client obtained in this iteration are consistent with the strategies obtained in the previous iteration.

[0124] (3) Lightweight strategy solving module: This module includes an aggregator strategy optimizer and a client strategy solver. The theoretical basis of this module is that the game equilibrium in the bidirectional incentive collaboration module exists and is unique.

[0125] Game equilibrium is defined as an existing and unique equilibrium where, when the aggregator and each client reach game equilibrium, for the aggregator, no other reward strategy can make the aggregator more efficient than the current reward strategy; and for each client, no other data contribution strategy can make the client more efficient than the current data contribution strategy.

[0126] In the t-th round of global iteration, the specific execution flow of the aggregator policy optimizer is as follows:

[0127] Step 1: Calculate parameters for each client: The calculation formula is:

[0128] Step 2: Let D1, D2, ..., D be the maximum amount of data that each client can contribute. KAccording to these clients The values ​​are sorted in descending order, that is, for the s-th value after sorting... i s i+1 One client,

[0129] Step 3: Divide the range of values ​​for the aggregator reward decision into:

[0130]

[0131] Step 4: Calculate the local optimum when the aggregator reward policy value lies within each interval. When the local optimal solution is:

[0132]

[0133] when In this case, the utility function design in the two-way incentive system module can guarantee the existence and uniqueness of a locally optimal reward strategy within this interval. Therefore, a lightweight solution method, Newton's iteration method, is adopted to calculate the locally optimal solution within this interval.

[0134] when At that time, local optimal reward in, For locally optimal reward, To be according to After sorting in descending order, the first one is ranked D in i , To be according to After sorting in descending order, the last one is ranked D in i , To be according to After sorting in descending order, the first one is ranked In To be according to After sorting in descending order, the last one is ranked In ν1 and v2 are balancing parameters in the aggregator utility function used to balance the aggregator's emphasis on global model accuracy and reward cost, φ tH For precision control parameters, For precision control parameters, w i For the i-th client, For a client collection, w j For the j-th client, ψ is the precision control parameter. Here, H represents the intermediate calculation parameters, and H represents the number of local iterations during federated learning training on each client. This is the set of clients selected in the t-th round of global iteration.

[0135] Step 5: Calculate the globally optimal reward strategy

[0136] In the t-th round of global iteration, the specific execution flow of the client policy solver is as follows:

[0137] Step 1: Determine τ t Is it 0? If it is 0, then it represents the optimal data contribution strategy for each client. Otherwise, proceed to step two.

[0138] Step 2: Perform this step for each client. Taking the i-th client as an example, determine τ. t Is it less than If it is less than, proceed to step three; otherwise, determine the optimal data contribution strategy for the i-th client.

[0139] Step 3: For clients entering Step 3, the optimal data contribution strategy is as follows:

[0140] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A dynamic selection and joint incentive optimization system for federated learning clients in edge computing, characterized in that, include: The module includes a dynamic quality perception screening module, a two-way incentive collaboration module, and a lightweight strategy solving module. The dynamic quality-aware filtering module is used to intelligently filter target value clients to participate in training based on the data quality of each client before the federated iteration. The bidirectional incentive collaboration module is used to solve for the optimal strategy combination based on the aggregator and the selected clients using the lightweight strategy solving module, and to optimize the decisions of the aggregator and clients based on the optimal strategy combination. Optimizing aggregator and client decisions based on the optimal strategy combination includes: Step 1: Set an upper limit on the number of game iterations. Randomly initialize the aggregator reward decision; Step 2: Each client calculates the optimal data contribution strategy within the current game iteration using the lightweight strategy solving module, based on the reward strategy initialized by the aggregator. Step 3: The aggregator calculates the optimal reward strategy within the current game iteration based on the optimal data contribution strategy of each client within the current game iteration using the lightweight strategy solving module. Step 4: Repeat steps 2 and 3 until the maximum number of iterations is reached. The iteration continues until the strategies of the aggregator and each client obtained in this iteration are consistent with the strategies obtained in the previous iteration, at which point the iteration stops and the optimized aggregator and client decisions are obtained. Each client calculates the optimal data contribution strategy based on the reward strategy initialized by the aggregator, using the lightweight strategy solving module, including: Step 1: Determine if the reward strategy after aggregator initialization is 0. If it is 0, the optimal data contribution strategy for each client is 0; otherwise, proceed to Step 2. Step 2: Determine whether the reward strategy after aggregator initialization is less than the judgment value. If it is less than, proceed to step three; otherwise, proceed to step three. Optimal data contribution strategy for each client ; Step 3: The reward strategy after aggregator initialization is less than the judgment value. For clients, the optimal data contribution strategy is ,in, For the optimal contribution strategy, The reward strategy after the aggregator is initialized. Indicates the first The actual data quality of each client in this round of global iteration The worst data quality that the aggregator can tolerate. Let be the cost coefficient for the i-th client. Let i be the total amount of data for the i-th client; The lightweight strategy solving module is used to calculate the optimal reward strategy of the aggregator and the optimal data contribution strategy of each client, and obtain the optimal strategy combination.

2. The federated learning client dynamic selection and joint incentive optimization system for edge computing as described in claim 1, characterized in that, The dynamic quality-aware screening module includes: a lightweight gradient difference analyzer and an online quality tracker; The lightweight gradient difference analyzer is used to filter out the data with the highest quality based on the historical quality information of each client. One client participates in federated learning; The online quality tracker is used to record the data quality information of the clients participating in this round of training, update the screening criteria for each client, and use it for the next round of client screening.

3. The federated learning client dynamic selection and joint incentive optimization system for edge computing as described in claim 2, characterized in that, Used to filter out the highest quality data based on historical quality information from each client. The number of clients participating in federated learning includes: The historical quality information of the clients is sorted in descending order, and the clients with the highest data quality are selected from the descending order. Each client participates in federated learning.

4. The federated learning client dynamic selection and joint incentive optimization system for edge computing as described in claim 2, characterized in that, Used to record the data quality information of clients participating in this round of training, update the selection criteria for each client, and for the next round of client selection, including: Updated up to the The cumulative number of times each client participates in federated learning during each round of global iteration; Based on the cumulative number of times participated in federated learning, updated up to the [number]th [date / period]. The average data quality of each client during the global iteration; Based on the average data quality, update the... CUCB-based data quality for each client in the round; Based on the quality of the CUCB-based data, the screening criteria for each client are updated.

5. The federated learning client dynamic selection and joint incentive optimization system for edge computing as described in claim 1, characterized in that, The aggregator calculates the optimal reward strategy for the current game iteration based on the optimal data contribution strategy of each client using a lightweight strategy solving module. Step 1: Calculate the parameters for each client. ; Step 2: Obtain the maximum amount of data that each client can contribute. , will the client according to The values ​​are sorted in descending order; Step 3: Based on the clients sorted in descending order, divide the range of values ​​for the aggregator reward decision; Step 4: Calculate the locally optimal reward for each range of values ​​in the aggregator reward strategy; Step 5: Based on the local optimal reward, obtain the global optimal reward strategy within the current game iteration.

6. The dynamic selection and joint incentive optimization system for federated learning clients for edge computing according to claim 5, characterized in that, The range of values ​​for the aggregator reward decision is divided into: ； when At that time, the locally optimal reward is: ； when At that time, local optimal reward ; when At that time, local optimal reward ,in, For locally optimal reward, To be according to After sorting in descending order, the first one is ranked In , To be according to After sorting in descending order, the last one is ranked In , To be according to After sorting in descending order, the first one is ranked In , To be according to After sorting in descending order, the last one is ranked In , , This is a balancing parameter in the aggregator utility function used to balance the aggregator's emphasis on global model accuracy and reward cost. For precision control parameters, For precision control parameters, For the i-th client, For client collection, For the j-th client, For precision control parameters, These are intermediate calculation parameters. This refers to the number of local iterations performed by each client during federated learning training. This is the set of clients selected in the t-th round of global iteration.

Images