Federal learning client dynamic screening and joint excitation optimization system oriented to edge computing

Abstract

The invention discloses an edge computing-oriented federated learning client dynamic screening and joint excitation optimization system. The system comprises a dynamic quality perception screening module, a bidirectional excitation collaboration module and a lightweight strategy solving module, the dynamic quality sensing and screening module is used for intelligently screening a target value client to participate in training according to the data quality of each client before federal iteration; the bidirectional incentive collaboration module is used for solving an optimal strategy combination by utilizing the lightweight strategy solving module based on the aggregator and the screened client, and optimizing the aggregator and client decision according to the optimal strategy combination; and the lightweight strategy solving module is used for calculating an optimal reward strategy of the aggregator and an optimal data contribution strategy of each client to obtain an optimal strategy combination. According to the method, excitation interaction between the aggregator and the client represented by the edge device in the edge scene is established, the global model precision is improved under the condition that the data quality of the client is unknown, and unfair excitation is avoided.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of edge computing and federated learning, and particularly relates to an edge computing-oriented federated learning client dynamic screening and joint incentive optimization system. BACKGROUND

[0002] Edge computing deploys computing, storage and network capabilities on the network edge close to the data source, aiming to meet the needs of low latency, high bandwidth, data localization processing and security privacy, and is widely used in real-time state monitoring of factory production line equipment, real-time traffic flow analysis of roadside units, personalized services of wearable devices, etc. In recent years, with the deep integration of Internet of Things, 5G / 6G and artificial intelligence, massive data is generated in these scenarios. Using these data for machine learning model training is the core of intelligent decision-making. This puts forward the requirements of ultra-low latency, high bandwidth and high reliability for data processing and intelligent decision-making. However, the traditional centralized cloud training mode cannot meet these needs due to long data transmission distance and network congestion. Specifically, in terms of data privacy protection, edge data often contains sensitive information such as factory core process parameters and personal biological features. With the increasing strictness of data privacy regulations, the traditional centralized cloud training poses a risk of raw data leakage due to large-scale data transmission. In terms of network bandwidth, uploading PB / EB level raw edge data, especially video, image and high-frequency sensor streams, to the cloud continuously consumes huge bandwidth resources, is costly, and is not feasible or highly inefficient in remote factories and mobile cellular network bandwidth limited scenarios. In terms of edge device resource utilization, modern edge devices such as embedded systems, industrial controllers, smartphones and roadside units have limited computing power, memory, storage space and battery life, but have certain computing capabilities (CPU / GPU / NPU). The centralized mode fails to effectively utilize these distributed computing potential. The above challenges seriously restrict the real-time performance, security and economy of edge intelligent applications, and new distributed machine learning paradigms are needed to break through the bottleneck.

[0003] Federated learning, as a distributed machine learning technology paradigm, provides an effective way to solve the above edge computing pain points. Its core innovation is that the original training data is retained on the local edge device. The training process is achieved by iteratively performing the following steps:

[0004] Model distribution: the aggregator deployed in the edge cloud or regional data center distributes the current global model to selected edge clients.

[0005] Local training: the client uses local private data and local computing resources to train based on the global model, and calculates model gradients, model weight differences and other update parameters.

[0006] Secure upload: Clients encrypt and upload only model update results (not raw data) to the central coordinator.

[0007] Secure aggregation: Aggregator securely aggregates multiple model update results received, generating a new global model.

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

[0009] The technical advantages of federated learning can overcome the difficulties of traditional centralized learning when applied in edge computing scenarios. In predictive maintenance of industrial Internet of Things, federated learning enables each device to train a fault prediction model locally and only share model updates. This protects core data of the factory, reduces data transmission across production lines or across factory areas, and utilizes the computing power of the device itself. In cooperative sensing and decision-making of Internet of Vehicles, federated learning enables roadside units to train a sensing model locally or contribute to local learning results, enabling rapid environmental awareness update and cooperative decision-making, and avoiding bandwidth explosion and delay caused by uploading of raw video / stream data. In real-time monitoring and time response of smart cities, federated learning enables cameras or edge servers to analyze video streams locally, train or update target detection or behavior recognition models, and only share key model updates or event summaries, thus protecting privacy and improving response speed.

[0010] In federated learning, the accuracy of the global model is usually related to the total amount and quality of the data of the client datasets participating in federated learning. The data quality refers to the difference between the data amount distribution of each label in the data of the client and the total label distribution of the data of all the clients participating in federated learning. The greater the difference, that is, the greater the non-identically distributed degree of the client dataset, the poorer the data quality, because the local models trained on different non-identically distributed datasets usually reduce the accuracy of the global model when aggregated. In common edge computing scenarios, due to the differences in time, space, and sensors for collecting data, the non-identically distributed degree of the datasets of various clients is highly different. Therefore, if the clients are allowed to participate in federated learning without screening, the accuracy of the global model is usually not improved, and even reduced. More importantly, the participation of low-quality clients will waste the limited power, computing period and network bandwidth of edge devices, and aggravate the invalid consumption of edge resources. However, the traditional federated learning framework usually assumes that the data quality of the clients is known to the aggregator, which is difficult to achieve in the actual application of federated learning, because the federated learning protects the privacy of the client data, and the aggregator cannot directly observe the original dataset of the client, so it cannot evaluate the data quality of the client, and then screen the clients participating in federated learning according to the data quality of the client. In addition, the training of the local model by the client will bring about the cost of communication energy consumption, computing energy consumption, etc. If the aggregator cannot give the client sufficient payment, the client will not actively contribute data to participate in federated learning, which will reduce the total amount of data for federated learning, and thus reduce the accuracy of the global model.

[0011] In summary, there are still the following three challenges in the design of the federated learning incentive mechanism with client selection: first, the data quality is unknown. In actual edge-side federated learning applications, due to the privacy protection mechanism of federated learning, the non-identically distributed degree of the client dataset cannot be directly evaluated, that is, the data quality is usually unknown to the aggregator, which makes the screening effect poor, causes resource waste and aggravates the accuracy loss; second, the evaluation of the accuracy of the global model is distorted. When designing the incentive mechanism to maximize the accuracy of the global model, an evaluation function of the accuracy of the global model usually needs to be designed. Due to the non-convexity of the loss function of the machine learning model, the preset accuracy proxy function (such as the logarithmic loss) cannot reflect the real improvement, resulting in inaccurate incentive allocation, and high-quality updates do not get enough return; third, the strategy solution is not feasible. When the aggregator and the client calculate the optimal strategy through heuristic search or reinforcement learning under information asymmetry, the optimality of the obtained strategy cannot be guaranteed, which further leads to the loss of the accuracy of the global model, and the edge-side aggregator lacks computing power, which cannot generate reliable strategies in real time, and the risk of system paralysis is high. SUMMARY

[0012] To solve the above technical problems, the application provides an edge computing-oriented federated learning client dynamic screening and joint incentive optimization system, establishes the incentive interaction between the aggregator and the client represented by the edge device in the edge scenario, improves the global model accuracy in the case of unknown client data quality, and avoids unfair incentives.

[0013] To achieve the above object, the application provides an edge computing-oriented federated learning client dynamic screening and joint incentive optimization system, which comprises a dynamic quality perception screening module, a bidirectional incentive cooperation module and a lightweight strategy solving module.

[0014] The dynamic quality perception screening module is used to intelligently screen target value clients participating in training according to the data quality of each client before federated iteration.

[0015] The bidirectional incentive cooperation module is used to solve the optimal strategy combination by using the lightweight strategy solving module based on the aggregator and the screened clients, and optimize the aggregator and client decision according to 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 perception screening module comprises a lightweight gradient difference analyzer and an online quality tracker.

[0018] The lightweight gradient difference analyzer is used to screen 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 the current round of training, update the screening criteria of each client, and be used for the next round of client screening.

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

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

[0022] Optionally, recording the data quality information of the clients participating in the current round of training, updating the screening criteria of each client, and being used for the next round of client screening comprises:

[0023] Updating the cumulative number of times each client participates in federated learning up to the tth global iteration;

[0024] updating an average data quality of each client as of a t-th round global iteration based on the accumulated number of participating in federated learning;

[0025] updating a CUCB-based data quality of each client in the t-th round based on the average data quality;

[0026] updating a screening criterion of each client based on the CUCB-based data quality.

[0027] Optionally, the optimizing the aggregator and the client decisions according to the optimal strategy combination comprises:

[0028] Step one, setting an upper limit iter of game iteration times max , and randomly initializing an aggregator reward strategy;

[0029] Step two, each client calculating an optimal data contribution strategy in a current game iteration through a lightweight strategy solving module according to the aggregator's initialized reward strategy;

[0030] Step three, the aggregator calculating an optimal reward strategy in the current game iteration through the lightweight strategy solving module according to the optimal data contribution strategy of each client in the current game iteration;

[0031] Step four, repeating steps two and three until the iteration times reach the upper limit iter max , or until the strategies of the aggregator and each client obtained in the current iteration are consistent with those obtained in the last iteration, then stopping the iteration and obtaining the optimized aggregator and client decisions.

[0032] Optionally, the each client calculating the optimal data contribution strategy according to the aggregator's initialized reward strategy through the lightweight strategy solving module comprises:

[0033] Step one, judging whether the aggregator's initialized reward strategy is 0, if yes, the optimal data contribution strategy of each client is 0, otherwise, entering step two;

[0034] Step two, judging whether the aggregator's initialized reward strategy is less than a judgment value , if yes, entering step three, otherwise, the optimal data contribution strategy of the i-th client is

[0035] Step three, the aggregator's initialized reward strategy is less than the judgment value , the optimal data contribution strategy of the client is , wherein, is the optimal contribution strategy, τ t is the aggregator's initialized reward strategy, represents the actual data quality of the i-th client in the current global iteration, λ is the worst data quality that the aggregator can tolerate, a i is the cost coefficient of the i-th client, D i is the total data amount of the i-th client.

[0036] Optionally, the aggregator calculates the optimal reward strategy in the current game iteration according to the optimal data contribution strategy of each client in the current game iteration through the lightweight strategy solving module, which includes:

[0037] Step one, calculate the parameters of each client

[0038] Step two, obtain the maximum data amount D1, D2, …, D K that each client can contribute; arrange the clients in descending order of the value of D

[0039] Step three, based on the clients arranged in descending order, divide the value range of the aggregator reward decision;

[0040] Step four, calculate the local optimal reward of the value range of the aggregator reward strategy;

[0041] Step five: based on the local optimal reward, obtain the global optimal reward strategy in the current game iteration.

[0042] Optionally, the value range of the aggregator reward decision is divided into:

[0043]

[0044] When , the local optimal reward is:

[0045]

[0046] When , the local optimal reward

[0047] When , the local optimal reward wherein, is the local optimal reward, is the D i in the first-ranked after being arranged in descending order of D , is the D i in the last-ranked after being arranged in descending order of D , is the D After descending arrangement, the first in is arranged in descending order After descending arrangement, the last in v1, v2 are balance parameters in the aggregator utility function for balancing the importance of the global model accuracy and reward cost for the aggregator, and phi tH is an accuracy control parameter, is an accuracy control parameter, w i is the i-th client, is the client set, w j is the j-th client, and psi is an accuracy control parameter, is an intermediate calculation parameter, and H is the local iteration number when each client locally executes federated learning training, is the client set selected in the t-th round of global iteration.

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

[0049] The present application realizes the following breakthrough effects in edge federated learning through the dynamic quality perception screening and the synergistic joint optimization mechanism of the two-way incentive coordination:

[0050] (1) The global model accuracy is significantly improved: high-quality clients are dynamically selected through gradient difference analysis, and the participation rate of low-quality clients is reduced, so that the model accuracy is improved under the same iteration round. The data quality and resource cost are included in the utility function of the incentive module, which drives the clients to contribute high-quality data, and the model convergence speed per unit data volume is improved.

[0051] (2) The edge resource consumption is greatly optimized: the CUCB mechanism of the quality perception module reduces invalid exploration, saves client computing energy consumption, and prolongs the life of battery devices. The lightweight strategy solution reduces the CPU occupancy rate of the aggregator, and adapts to resource-limited edge nodes.

[0052] (3) Privacy protection and fairness guarantee: the data quality is inferred through model gradient difference throughout the whole process (without access to raw data), which meets the privacy protection requirements. The incentive mechanism realizes pay-per-quality, which solves the fairness problem of "low compensation for high-quality data" in traditional solutions.

[0053] (4) System robustness is enhanced: the uniqueness of the Steinberg game equilibrium ensures the stability of the strategy convergence, which reduces the strategy shock compared with traditional reinforcement learning methods. It supports dynamic edge environment (devices randomly offline, network fluctuation), and reduces the training task interruption rate. BRIEF DESCRIPTION OF DRAWINGS

[0054] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The embodiments of this application, and of the prior art, will be described and explained with additional specificity and detail by the accompanying drawings:

[0055] Figure 1 is a federated learning client dynamic screening and joint incentive optimization system structure diagram for edge computing. DETAILED DESCRIPTION

[0056] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

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

[0058] In view of the deficiencies of the traditional federated learning paradigm in the application of edge computing scenarios, the present application constructs the mapping relationship between the data distribution of each client dataset and the global model convergence performance according to the real-time local model update of the client, the global model convergence status analysis and other related information, under the condition that the specific situation of the client dataset is unknown, and designs an incentive mechanism to encourage the client to actively participate in federated learning, so that the behavior of the aggregator and the client changes towards the direction of global model convergence, thereby realizing the participation of the client in federated learning under the premise of unknown client data quality, maximizing the global model precision. The present application selects the client with high data quality to participate in federated learning, and encourages the client to contribute more data to participate in federated learning, solving the challenge of difficult optimization of global model precision under the condition of unknown client data quality.

[0059] The present embodiment proposes a federated learning client dynamic screening and joint incentive optimization system for edge computing as shown in Figure 1 , specifically comprising a dynamic quality perception screening module, a two-way incentive cooperation module and a lightweight strategy solving module.

[0060] The dynamic quality perception screening module is used to intelligently screen target value clients for participation in training according to the data quality of each client before federated iteration;

[0061] The two-way incentive cooperation module is used to solve the optimal strategy combination by using the lightweight strategy solving module based on the aggregator and the screened clients, and to optimize the aggregator and client decision-making according to the optimal strategy combination.

[0062] A lightweight strategy solving module is configured to calculate an optimal reward strategy of the aggregator and an optimal data contribution strategy of each client, and obtain an optimal strategy combination.

[0063] Specifically, the dynamic quality-aware screening module: before each round of federated iteration, this module needs to intelligently screen high-value clients to participate in training according to the data quality of each client. First, a lightweight gradient difference analyzer is constructed to dynamically generate a data quality evaluation value based on the difference amplitude of the historical local gradient model and the global gradient model of the client. Then, an online quality tracker is designed to record the actual quality performance of the client after participation, and to establish a client quality history archive. Through the above methods, a privacy-protected evaluation is realized, which does not need to access the original data throughout the process and only relies on model update gradient.

[0064] (2) Two-way incentive coordination module: in order to make the behaviors of the aggregator and the clients contribute to improving the accuracy of the global model, this module synchronously optimizes the aggregator and client decisions through a reward mechanism based on Steinberg game and federated learning convergence analysis. This module adopts a hierarchical decision framework, with the aggregator as the leader to regulate the reward budget to maximize the global accuracy, and the client as the follower to adjust the data contribution amount according to the reward. A precision-cost joint utility function is designed, with the aggregator utility being a global model accuracy evaluation function based on federated learning convergence analysis minus reward expenditure, and the client utility being reward income minus local resource consumption cost. Through the establishment of a game-driven convergence mechanism, system equilibrium is achieved by using two-way strategy adjustment.

[0065] (3) Lightweight strategy solving module: the goal of this module is to efficiently calculate the optimal strategy combination, which consists of a client strategy solver and an aggregator strategy optimizer. The client strategy solver directly calculates the optimal data contribution amount of each client by gradient backpropagation, and the aggregator strategy optimizer solves the reward allocation scheme by using an efficient numerical iteration algorithm. The computational complexity of the solving algorithm is controlled at O(n), which is suitable for edge aggregator computing power.

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

[0067] The lightweight gradient difference analyzer is configured to screen K clients with the highest data quality to participate in federated learning based on historical quality information of each client;

[0068] The online quality tracker is configured to record data quality information of the clients participating in the current round of training, update the screening criteria of each client, and be used for screening of the clients in the next round.

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

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

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

[0072] The cumulative number of times each client participates in federated learning as of the tth global iteration is updated;

[0073] Based on the cumulative number of times each client participates in federated learning, the average data quality of each client as of the tth global iteration is updated;

[0074] Based on the average data quality, the UCB-based data quality of each client in the tth round is updated;

[0075] Based on the UCB-based data quality, the selection criteria of each client are updated.

[0076] Further, optimizing the aggregator and client decisions according to the optimal strategy combination includes:

[0077] Step one, set the upper limit of the number of game iterations iter max , randomly initialize the aggregator reward strategy;

[0078] Step two, each client calculates the optimal data contribution strategy in the current game iteration through the lightweight strategy solving module according to the aggregator's initialized reward strategy;

[0079] Step three, the aggregator calculates the optimal reward strategy in the current game iteration through the lightweight strategy solving module according to the optimal data contribution strategy of each client in the current game iteration;

[0080] Step four, repeat steps two and three until the number of iterations reaches the upper limit iter max , or until the strategies of the aggregator and each client obtained in this iteration are consistent with those obtained in the last iteration, then stop the iteration and obtain the optimized aggregator and client decisions.

[0081] Further, each client calculates the optimal data contribution strategy through the lightweight strategy solving module according to the aggregator's initialized reward strategy, including:

[0082] Step one, determine whether the aggregator's initialized reward strategy is 0, if it is 0, the optimal data contribution strategy of each client is 0, otherwise go to step two;

[0083] Step two, determine whether the aggregator's initialized reward strategy is less than the determination value If less, go to step three, otherwise the optimal data contribution strategy of the i-th client

[0084] Step three, the reward strategy of the aggregator after initialization is less than The optimal data contribution strategy of the i-th client is Wherein, is the optimal contribution strategy, τ t is the reward strategy of the aggregator after initialization, represents the actual data quality of the i-th client in the current round of global iteration, λ is the worst data quality that the aggregator can tolerate, a i is the cost coefficient of the i-th client, D i is the total data amount of the i-th client.

[0085] Further, the aggregator calculates the optimal reward strategy in the current game iteration according to the optimal data contribution strategy of each client in the current game iteration through the lightweight strategy solving module, including:

[0086] Step one, calculate the parameters of each client

[0087] Step two, obtain the maximum amount of data D1, D2, …, D K that each client can contribute; arrange the clients in descending order according to the value of

[0088] Step three, based on the clients arranged in descending order, divide the value range of the reward decision of the aggregator;

[0089] Step four, calculate the local optimal reward of the value range of the reward strategy of the aggregator;

[0090] Step five: based on the local optimal reward, obtain the global optimal reward strategy in the current game iteration.

[0091] The present embodiment is described taking the automobile manufacturing supply chain quality prediction federated learning as an example:

[0092] A certain vehicle manufacturer wants to build a high-precision parts defect prediction model to identify high-risk batches before entering the assembly line and reduce rework costs. The model needs real-time production data from multiple key parts suppliers (such as A company provides engine block, B company provides electronic control system, C company provides brake parts) to participate in training. However, training the model faces multiple challenges: data sovereignty and privacy, each supplier's production data (such as sensor readings, process parameters, quality inspection records) is their core competitiveness and business secrets, suppliers strongly oppose sharing raw data directly with the vehicle manufacturer or other suppliers; data quality heterogeneity, each parts supplier's data collection method, scene is different, the quality of collected data is uneven. Incentive needs, participating in federated learning will consume the supplier's own resources (computing power, electricity, network bandwidth, engineer maintenance time), the vehicle manufacturer needs to provide sufficient economic compensation or business benefits, otherwise there is no motivation to contribute data or seriously train. "Free rider" risk, if there is no reasonable incentive, high-quality supplier A may reduce contribution (because of high cost), and low-quality supplier B may try to contribute low-value data to cheat rewards.

[0093] By applying the embodiment scheme, a global defect prediction model aggregator is deployed on the edge cloud of the vehicle manufacturer, and each supplier's local factory edge node is used as a client to store private data and train a local defect prediction model. To collaboratively train a global parts defect prediction model, in each global iteration: the aggregator first selects the top K clients with the highest data quality using the lightweight gradient difference analyzer in the dynamic quality perception screening module based on historical client data quality information. The selected clients and the aggregator enter a two-way incentive collaboration module together to form a Stakelberg game, and the optimal reward calculation strategy of the aggregator and the optimal data contribution strategy of each client are iteratively calculated using the lightweight strategy solver module until the game reaches equilibrium. Then, the aggregator sends the current global defect prediction model parameters to the selected clients, each client trains a local defect prediction model based on its optimal data contribution strategy and global model parameters, and uploads the local model parameters to the aggregator after training. The aggregator completes global model aggregation and pays rewards to each client participating in federated learning. Finally, the online quality tracker in the dynamic quality perception screening module records the actual data quality of the clients participating in the current federated learning and updates the quality parameters of each client for the next global iteration.

[0094] (1) Dynamic quality-aware screening module: This module consists of two parts, lightweight gradient difference analyzer and online quality tracker. When screening the clients to participate in federated learning, the lightweight gradient difference analyzer selects the top K clients with the highest data quality based on the historical information of the data quality of each client when participating in federated learning before. After the selected clients complete the current round of federated learning, the online quality tracker records the data quality information of these clients in this round, and updates the screening criteria of each client for the next round of client screening.

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

[0096] Step one: arrange each client in descending order according to its CUCB-based data quality Assuming a total of N clients, where CUCB-based data quality is data quality based on confidence upper bound, and CUCB algorithm is a combination of confidence upper bound algorithm. This CUCB-based data quality is a key intermediate variable in the execution of this algorithm.

[0097] Step two: select the top K clients from the descendingly arranged clients to participate in federated learning.

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

[0099] Step one: update the cumulative number of times each client participates in federated learning up to the t-th round of global iteration The update formula is as follows:

[0100]

[0101] Where, is an indicator function, which outputs 1 when the function input is true, and 0 otherwise.

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

[0103]

[0104] Where, represents the actual data quality of the i-th client in the current 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 client participating in federated learning through the following formula:

[0105]

[0106] wherein, denote the global model and the gradient of the i-th client local model, respectively, g max denotes the maximum expected label loss, denotes the bulldozer distance of the i-th client dataset.

[0107] Step three: update the UCB-based data quality of each client in the t-th round The update formula is:

[0108]

[0109] The CUCB-based data quality will be used as the basis for the lightweight gradient difference analyzer to select clients. This data quality ensures that clients with fewer selection times have more opportunities to be selected in future global iterations, thereby avoiding missing clients with high data quality, i.e., optimism in the face of uncertainty.

[0110] (2) Two-way incentive coordination module: After selecting the clients participating in federated learning in the t-th round of global iteration, this module will aggregate the leader and the selected client as the follower to form a two-stage Steinberg game incentive mechanism. After the aggregator gives its reward strategy, the selected client determines its data contribution strategy according to the reward strategy, and repeats the process until the game equilibrium.

[0111] The definition of game equilibrium is that the leader and the follower cannot further improve their own benefits by changing their own decisions without changing the strategies of other participants. Specifically, in the federated learning system, the aggregator cannot improve its own utility (global model accuracy - reward cost) by changing its own reward decision only, because the aggregator will change the data contribution strategy of the client as the follower once it changes its own reward decision. Similarly, the client cannot improve its own utility (reward benefit - resource cost) by changing its own reward decision only.

[0112] The utility function of the t-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 the 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 affect the data contribution strategy to maximize its utility function value.

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

[0117]

[0118] where the first term is the reward income of the client and the second term is the resource consumption cost. The client influences the aggregator's reward policy τ by adjusting its data contribution strategy t , and then maximizes its utility function value.

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

[0120] Step one: set the upper limit of the number of iterations iter max , and randomly initialize the aggregator's reward decision τ t .

[0121] Step two: each client calculates the optimal data contribution strategy through the lightweight strategy solving module according to the aggregator's reward policy.

[0122] Step three: the aggregator calculates the optimal reward strategy through the lightweight strategy solving module according to the data contribution strategies of each client.

[0123] Step four: repeat steps two and three until the number of iterations reaches the upper limit iter max , or until the strategies of the aggregator and each client obtained in this iteration are consistent with those obtained in the last 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 two-way incentive coordination module exists and is unique.

[0125] The definition of the existence and uniqueness of the game equilibrium is: when the aggregator and each client reach the game equilibrium, for the aggregator, there is no other reward strategy that can make the utility of the aggregator higher than the current reward strategy; for each client, there is no other data contribution strategy that can make the utility of the client higher than the current data contribution strategy.

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

[0127] Step one: calculate the parameters for each client: The calculation formula is:

[0128] Step two: record the maximum amount of data that each client can contribute as D1, D2, …, D KThe clients are ranked in descending order of the value of , i.e., for the s i th and s i+1 th clients in the sorted order,

[0129] Step three: divide the value range of the aggregator reward decision into:

[0130]

[0131] Step four: calculate the local optimal solution when the value of the aggregator reward strategy is in each interval. When , the local optimal solution is:

[0132]

[0133] When , the utility function design in the two-way incentive system module can guarantee the existence and uniqueness of the local optimal reward strategy in this interval, so the lightweight solving method, Newton iteration method, is used to calculate the local optimal solution in this interval

[0134] When , the local optimal reward where, is the local optimal reward, is the D in the first place after ranking in descending order, is the D i in the last place after ranking in descending order, is the D in the first place after ranking in descending order, is the D in the last place after ranking in descending order, is the D in the first place after ranking in descending order, is the D in the last place after ranking in descending order, ν1, v2 are balance parameters in the aggregator utility function for balancing the importance of the global model accuracy and reward cost for the aggregator, tH is the accuracy control parameter, is the accuracy control parameter, w i is the i th client, is the client set, w j is the j th client, ψ is the accuracy control parameter, is an intermediate calculation parameter, H is the local iteration number when each client performs federated learning training locally, the client set screened out in the tth global iteration.

[0135] Step five: calculate the global optimal reward strategy

[0136] In the tth global iteration, the specific execution process of the client strategy solver is as follows:

[0137] Step one: judge whether τ t is 0, if yes, the optimal data contribution strategy of each client is Otherwise, go to step two.

[0138] Step two: execute this step for each client. Take the ith client as an example, judge whether τ t is less than If yes, go to step three, otherwise the optimal data contribution strategy of the ith client is

[0139] Step three: for the client entering step three, the optimal data contribution strategy is

[0140] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited to this, any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection 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. 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 select the K clients with the highest data quality to participate in federated learning based on the historical quality information of each client. 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, The method used to select the K clients with the highest data quality based on the historical quality information of each client to participate in federated learning includes: 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.

4. The dynamic selection and joint incentive optimization system for federated learning clients oriented towards edge computing according to 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: Update the cumulative number of times each client has participated in federated learning up to the t-th round of global iteration; 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; Based on the average data quality, update the CUCB-based data quality of each client in round t; 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, Optimizing aggregator and client decisions based on the optimal strategy combination includes: Step 1: Set an upper limit on the number of game iterations (iter) max 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 through the lightweight strategy solving module. 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.

6. The dynamic selection and joint incentive optimization system for federated learning clients for edge computing according to claim 5, characterized in that, 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, determine the optimal data contribution strategy for the i-th 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, 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.

7. The dynamic selection and joint incentive optimization system for federated learning clients for edge computing according to claim 6, 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, D1, D2, ..., D. K , 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.

8. The federated learning client dynamic selection and joint incentive optimization system for edge computing as described in claim 7, 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 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.

Images