Autonomous adjustment and data balance federated learning system and method based on incentive mechanism

Abstract

The invention belongs to the technical field of federated learning, and particularly relates to an autonomous adjustment and data balance federated learning system and method based on an incentive mechanism. The system comprises a client screening module for global heterogeneity perception, which is used for realizing dynamic evaluation and screening control of client data quality through a global label distribution offset perception mechanism; the server-side utility evaluation and reverse auction module is used for quantifying the actual contribution value of the client to global model training and realizing the distribution of excitation resources through an auction mechanism; and the client participation rate self-adaptive adjustment module is used for guiding the client to autonomously optimize the participation frequency according to the historical income and cost of the client based on a utility function so as to realize a long-term participation behavior. The method has the characteristics that an incentive mechanism is taken as a core, high-value participants are dynamically identified in combination with data characteristics and participation behaviors of a modeling client, and the system stability and the model performance are improved through a verifiable return strategy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of federated learning, and particularly relates to a federated learning system and method based on incentive mechanism and self-adjustment and data balancing. BACKGROUND

[0002] In recent years, with the implementation of the Data Security Law, the compliance circulation of data and the protection of security information have become key issues in the development of information technology. To meet the requirement of "data usability and invisibility", security information computing technology has developed rapidly. Among them, federated learning (FL) as an important security information computing paradigm, due to its characteristics of realizing multi-party joint modeling without sharing raw data, has attracted widespread attention in data-sensitive scenarios such as medical care, finance and industrial internet.

[0003] Federated learning coordinates multiple clients to collaboratively train a global model through a parameter server architecture. Each client updates the model parameters locally using private data and uploads them to the server for aggregation. After multiple iterations, the global model is obtained. However, in actual deployment, federated learning systems face the following three key challenges.

[0004] Firstly, data heterogeneity is significant. Due to differences in data collection environment, business type and distribution characteristics, different clients often have non-independent and identically distributed (non-IID) data, which manifests as a shift in label or feature distribution. This heterogeneity can significantly weaken the convergence efficiency and generalization performance of the global model, leading to unstable training process and even convergence failure.

[0005] Secondly, there is high uncertainty in client availability. In the Internet of Things and edge computing environment, client devices are affected by factors such as computing power, battery life, network connectivity, etc., showing dynamic volatility in participation status. This makes it difficult for some clients with high-quality data or computing power to participate stably in training, affecting model performance and training progress.

[0006] Finally, there is a lack of effective incentive mechanism. Federated learning requires clients to contribute local resources for training and communication. Without sufficient rewards, clients lack the motivation to participate continuously. This lack of incentives will lead to a decline in system participation, reducing overall stability and learning effectiveness.

[0007] Incentive mechanism in federated learning not only relates to the willingness of clients to participate, but also is a key means to address data heterogeneity and availability fluctuations. Reasonable incentive strategies can help attract high-quality and stable clients to participate continuously, thereby improving the convergence efficiency and generalization ability of the global model.

[0008] Firstly, in terms of data heterogeneity, most current methods still take the amount of data uploaded by the client as the main incentive basis, ignoring the quality of the data itself, the label distribution, and the actual contribution to the model training effect. This coarse-grained evaluation method may lead to low-quality or highly biased data being given too much reward weight, thereby interfering with the convergence process of the global model and even causing performance degradation. In addition, the lack of modeling means for data representativeness and diversity makes the model prone to overfitting to certain specific data patterns, reducing the generalization ability.

[0009] Secondly, in terms of availability modeling, existing mechanisms generally fail to fully consider the dynamic changes in the client training state. For example, some devices may frequently drop out, respond slowly, or even quit training midway due to limitations in computing power, network conditions, or energy consumption. Some clients also exhibit speculative behavior, participating only when high rewards are available. These problems not only disrupt training stability but also lead to server resource waste and decreased scheduling efficiency. Current systems lack mechanisms to constrain and punish such behavior, making it difficult to ensure the reliability and continuity of the participation process.

[0010] Finally, in terms of the fairness and credibility of the incentive mechanism, the client reward mechanism usually relies on centralized calculation or closed processes, lacking transparent and open means of income evaluation and verification, making it difficult to establish trust between clients and the system. In addition, current incentive strategies often fail to fully consider the participation costs of clients, including training time, computing power consumption, and communication costs, and do not provide differentiated or long-term incentives. This design may leave high-investment, high-value clients in a state of inadequate incentives for a long time, thereby weakening their enthusiasm for sustained participation and reducing the overall system stability.

[0011] Therefore, there is an urgent need for a new federated learning method that focuses on incentive mechanisms, jointly models the data characteristics and participation behavior of clients, dynamically identifies high-value participants, and improves system stability and model performance through verifiable reward strategies. SUMMARY

[0012] The present application is to overcome the problems of existing federated learning methods, which have significant data heterogeneity, high uncertainty in client availability, and lack of effective incentive mechanisms. It provides a self-regulating and data-balancing federated learning system and method based on incentive mechanisms, which focuses on incentive mechanisms, jointly models the data characteristics and participation behavior of clients, dynamically identifies high-value participants, and improves system stability and model performance through verifiable reward strategies.

[0013] To achieve the above invention purposes, the present application adopts the following technical solutions:

[0014] The self-regulating and data-balancing federated learning system based on incentive mechanisms comprises:

[0015] a client screening module for global heterogeneity perception, for realizing dynamic evaluation and screening control of client data quality through a global label distribution deviation perception mechanism;

[0016] a server-side utility evaluation and reverse auction module for quantifying the actual contribution value of clients to global model training, and realizing allocation of incentive resources through an auction mechanism;

[0017] a client participation rate adaptive adjustment module for guiding clients to autonomously optimize participation frequency based on historical income and cost, and realizing long-term participation behavior.

[0018] Preferably, the client screening module for global heterogeneity perception specifically comprises the following processes:

[0019] calculating an individual label distribution deviation degree: calculating the non-uniformity degree of the corresponding local label distribution for each client, defined as an individual label distribution deviation degree PCDI, for measuring the deviation degree of the local data of the client relative to the ideal IID distribution:

[0020]

[0021] where λ(i, l) represents the proportion of label l in client i, is an ideal uniform distribution;

[0022] collecting an online client set and sorting: before starting each round of training, the system collects the current available online client set and sorts the clients in the set C t from small to large according to the label heterogeneity indicator PCDI i , and C is the universal set of the client set;

[0023] constructing an initial client set: preferentially selecting the first ζ·|C t | clients with low data heterogeneity to form an initial participation set S t , and ζ represents a proportion parameter of the low-heterogeneity reference clients;

[0024] calculating an initial global label distribution indicator: for the current initial participation set S t , calculating the corresponding weighted global label distribution center, i.e., a global label uniformity indicator GCDI, for measuring the heterogeneity level of the current participating clients as a whole:

[0025]

[0026] where m j is the data set size of client j; λ for correcting the bias of model aggregation weights j Normalizing the distribution of local dataset for client j, for evaluating the degree of label bias of the overall dataset formed by the current set of clients

[0027] Joining the remaining clients to the set and judging the constraint condition: the server tries to join the remaining clients to S one by one in the order t , and recalculates the new GCDI as GCDI S ′ ; judge whether the relative growth of global heterogeneity index after the new client is joined meets the following condition:

[0028]

[0029] τ=1+(1+0.01·t)·β

[0030] Where t represents the current training round, and β is the adjustment factor

[0031] Eliminate clients that do not meet the condition: if the global heterogeneity growth does not exceed the threshold τ, accept the corresponding client to join the new set S t ; otherwise, exclude the corresponding client from the current round of screening.

[0032] As a preferred, the server-side utility evaluation and reverse auction module specifically includes the following processes:

[0033] Resume client evaluation index: the server collects the data size m i , historical participation frequency f i , and label heterogeneity index PCDI i of each client, and η is the weight balancing factor.

[0034] Calculate the comprehensive contribution index: establish a multi-dimensional index to evaluate the comprehensive contribution v i of each client:

[0035]

[0036] Client submits bid: clients with participation intention provide corresponding bid b i to the server.

[0037] Server calculates the cost performance and sorts: the server calculates the cost performance index rank i based on the client contribution index and bid:

[0038]

[0039] According to rank iSort from high to low, select several clients for training, the last client that meets the incentive condition under the limited budget B as the auction boundary value, recorded as rank s ;

[0040] Distribute aggregation weight and incentive: the server calculates the aggregation weight p i for each client

[0041]

[0042] At the same time, according to the price-performance ratio ranking and the auction boundary value, the final incentive amount is determined:

[0043]

[0044] Where, r i is the reward allocated by the server to the client i.

[0045] As a preferred, the client participation rate adaptive adjustment module specifically includes the following processes:

[0046] Construct the client utility function: after each round of training, the client constructs the expected utility function u i according to the training revenue and resource consumption as follows:

[0047]

[0048] Where, a i is the actual participation rate of the client, Δa i is the participation rate update amplitude of the client; c i is the fixed participation cost of each round of training, including communication and training cost; k i represents the additional cost paid for adjusting the participation rate;

[0049] Solve the optimal participation rate increment: the client aims to maximize its own utility, and solves the optimal participation rate adjustment amount Δa i *, to get the analytical solution:

[0050]

[0051] Calculate the authenticity constraint value: introduce authenticity constraint to ensure that the price-performance ratio of the client after adjustment is not lower than the original value, and the specific formula is as follows:

[0052]

[0053] Where, Δv i is the updated server-perceived client utility, and Δb i is the updated client bid. According to the constraint, the participation rate change is revised: represents the maximum participation rate adjustment value allowed under the premise of maintaining incentive authenticity; if then the Δa i should be clipped to meet the authenticity constraint:

[0054]

[0055] Update the client participation rate and modify the bid: the final client participation rate update formula is obtained as:

[0056] a i ←a i +Δa i

[0057] b i ←c i +k i ·(Δa i ) 2

[0059] The application also provides an autonomous adjustment and data balancing federated learning method based on an incentive mechanism, comprising the following steps:

[0060] S1, initializing the system, issuing a training task and calculating a label distribution heterogeneity index PCDI of the client i ;

[0061] S2, identifying a current online client set C t before starting each round of training;

[0062] S3, screening the top ζ proportion of clients in ascending order of PCDI to form an initial participation set S t ;

[0063] S4, calculating an initial global label balance index GCDI based on S t ;

[0064] S5, screening client participation based on a GCDI growth threshold mechanism;

[0065] S6, completing initial benchmark training client set construction;

[0066] S7, obtaining a client dataset size m i , a historical participation frequency f i index;

[0067] S8, calculating a comprehensive contribution degree v i of the client by comprehensively considering multi-dimensional contribution indexes;

[0068] S9, the client submits a bid b i of this round;

[0069] S10, the server calculates the cost-to-benefit ratio Determine the final selected set through the reverse auction mechanism, and generate the auction boundary value rank s ;

[0070] S11, assign the aggregation weight p to the selected client i And the training reward r i ;

[0071] S12, the client performs this round of training and uploads the model update;

[0072] S13, the client obtains the current auction boundary value rank s , and calculates the utility function based on the historical revenue and cost to update the participation rate;

[0073] S14, determine the optimal participation rate adjustment amount and modify the bid according to the new cost under the authenticity constraint;

[0074] S15, the server aggregates the global model;

[0075] S16, enter the next iteration, repeat steps S2 to S15 until the global model converges.

[0076] Compared with the prior art, the present application has the following advantages: (1) the present application realizes effective perception and dynamic control of client data heterogeneity, promotes data balance in global model training, and improves the stability and generalization ability of the model; (2) the present application constructs a fair and transparent incentive distribution mechanism, reasonably schedules client resources, and improves the overall training efficiency and diversity support of the system; (3) the present application dynamically adjusts the participation rate of the client, encourages high-value clients to participate in training continuously and stably, and enhances the robustness and long-term sustainability of the federated learning system. BRIEF DESCRIPTION OF DRAWINGS

[0077] Figure 1 A flowchart of the self-regulating and data-balancing federated learning method based on the incentive mechanism in the present application;

[0078] Figure 2 A framework diagram of the self-regulating and data-balancing federated learning system based on the incentive mechanism in the present application;

[0079] Figure 3 An interaction flowchart of the client and the server in the system of the present application. DETAILED DESCRIPTION

[0080] In order to more clearly illustrate the embodiments of the present application, the specific embodiments of the present application will be described below with reference to the drawings. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained from these drawings, and other embodiments can be obtained by those skilled in the art without creative labor.

[0081] The present application aims at the problems of resource allocation opacity, insufficient incentive, serious data heterogeneity and unstable client participation in the federated learning system in the edge computing environment, and provides a self-regulating and data-balancing federated learning system based on incentive mechanism. By introducing a client adaptive participation strategy, efficient resource allocation is realized, and the overall system performance is improved. The system specifically comprises:

[0082] A client screening module with global heterogeneity perception:

[0083] This module measures the non-equilibrium degree of the local label distribution of the client (PCDI) to preferentially select the client with more balanced label distribution to participate in training; meanwhile, a global label distribution index (GCDI) is designed to dynamically control the overall label offset of the selected client set. By dynamically adjusting the client screening standard through the round-adaptive tolerance threshold, this module ensures that the data distribution of the client tends to be balanced in each round of training, thereby improving the generalization performance and training stability of the model;

[0084] A server-side utility evaluation and reverse auction module:

[0085] After the preliminary screening of the client, this module further evaluates the training contribution potential of each client and guides it to submit the expected reward (bid). The server combines the historical behavior and bid of the client to construct a cost-performance ratio ranking and determine its participation qualification. Finally, according to the ranking and value index, the incentive and aggregation weight are allocated to ensure that the client return matches the contribution, thereby improving the resource utilization efficiency and stimulating the participation enthusiasm of high-quality clients;

[0086] A client participation rate adaptive adjustment module:

[0087] This module allows the client to autonomously adjust the participation frequency based on its historical income and participation cost. By constructing a utility function and solving the optimal strategy, the client can evolve its behavior round by round to improve the participation income while controlling the system overhead. This mechanism does not depend on central scheduling, enhancing the robustness and sustainable development capability of the system.

[0088] The federated learning method proposed in the present application systematically addresses the core problems of data heterogeneity, availability fluctuation and incentive fairness by introducing an incentive mechanism and client behavior modeling. Therefore, the following three collaborative modules are preferably constructed to realize the linkage optimization of data selection, incentive allocation and participation regulation:

[0089] The client screening module with global heterogeneity perception, the server-side utility evaluation and reverse auction module, and the client participation rate adaptive adjustment module.

[0090] Specifically, the client screening module with global heterogeneity perception specifically includes the following processes:

[0091] Calculate the individual label distribution deviation degree: calculate the non-equilibrium degree of the corresponding local label distribution for each client, defined as the individual label distribution deviation degree PCDI, which is used to measure the deviation degree of the local data of the client relative to the ideal IID distribution:

[0092]

[0093] Where λ(i, l) represents the proportion of label l in client i, is an ideal uniform distribution.

[0094] Collect and sort online client set: before starting each round of training, the system collects the current available online client set and sorts the clients in the set C t by the label heterogeneity indicator PCDI i from small to large, and C is the full set of client sets.

[0095] Construct an initial client set: preferentially select the first ζ·|C t | clients with low data heterogeneity to form an initial participation set S t , and ζ represents the proportion parameter of the low-heterogeneity reference clients.

[0096] Calculate the initial global label distribution indicator: for the current initial participation set S t , calculate the corresponding weighted global label distribution center, i.e., the global label balance indicator GCDI, which is used to measure the heterogeneity level of the current participating clients as a whole:

[0097]

[0098] Where m j is the data set size of client j. is used to correct the deviation of the model aggregation weight; λ j is the normalized distribution of the local data set of client j, which is used to evaluate the label deviation degree of the overall data set formed by the current client set.

[0099] Add the remaining clients to the set and judge the constraint condition: the server tries to add the remaining clients to S t in order, and recalculate the new GCDI, denoted as GCDI S ′; judge whether the relative growth of global heterogeneity index after the new client joins meets the following conditions:

[0100]

[0101] τ = 1 + (1 + 0.01·t)·β

[0102] Where t represents the current training round, and β is the adjustment factor.

[0103] Eliminate clients that do not meet the conditions: if the global heterogeneity growth does not exceed the threshold τ, accept the corresponding client to join the new set S t ; Otherwise, to avoid label drift diffusion, the corresponding client is excluded from the current round of screening. This process ensures that the training client set maintains overall balance and stability at the data level.

[0104] Further, the server-side utility evaluation and reverse auction module is used to quantify the actual contribution value of the client to the global model training, and to realize the fair and efficient allocation of incentive resources through the auction mechanism, including the following processes:

[0105] Resume client evaluation indicators: the server collects the data size m i , historical participation frequency f i , and label heterogeneity indicator PCDI i of each client, and η is the weight balance factor.

[0106] Calculate the comprehensive contribution index: establish a multi-dimensional index to evaluate the comprehensive contribution v i of each client:

[0107]

[0108] Client submits bid: the client with participation intention provides the corresponding bid b i to the server.

[0109] Server calculates the cost performance and sorts: the server side calculates the cost performance index rank i according to the client contribution index and bid:

[0110]

[0111] According to the rank i from high to low, select several clients for training, and the last client that meets the incentive condition under the limited budget B is recorded as rank s .

[0112] Allocate aggregate weights and incentives: to ensure reasonableness and fairness, the server calculates the aggregate weight p i:

[0113]

[0114] Meanwhile, the final incentive amount is determined according to the cost performance ranking and the auction boundary value:

[0115]

[0116] The mechanism takes into account the actual contribution of the client and the bid information, and improves the resource utilization and the incentive transparency.

[0117] Further, to cope with the fluctuation problem of the client participation behavior, the client participation rate adaptive adjustment module is designed to solve the problem of unstable client participation. The module guides the client to optimize the participation frequency based on the utility function according to the historical income and cost of the client, realizes rational and continuous long-term participation behavior, and specifically includes the following processes:

[0118] Constructing the client utility function: after each round of training, the client constructs the expected utility function u based on the training income and resource consumption, wherein the income part is the reward function of the server auction mechanism, and the cost part is established as a quadratic term of the fixed cost and the participation rate variation. i As follows:

[0119]

[0120] Wherein, a i is the actual participation rate of the client, Δa i is the participation rate update amplitude of the client; c i is the fixed participation cost of each round of training, including communication and training cost; k i represents the additional cost paid for adjusting the participation rate, which is a linearly increasing function of the participation rate itself, to reflect its marginal increasing property, so as to guide the client to achieve adaptive balance between income and cost;

[0121] Solving the optimal participation rate increment: the client takes maximizing its own utility as the goal, solves the optimal participation rate adjustment amount Δa i *, and obtains the analytical solution:

[0122]

[0123] Calculating the authenticity constraint value: to prevent the client from manipulating the ranking through non-authentic bidding to obtain excessive rewards, the authenticity constraint is introduced to ensure that the cost performance of the client after adjustment is not lower than the original value, and the specific formula is as follows:

[0124]

[0125] Wherein, Δv iFor the updated server-perceived client utility, Δb i For the updated client bid. According to the constraint, the participation rate change is revised: represents the maximum participation rate adjustment value allowed under the premise of maintaining incentive authenticity; if Δa i should be clipped and revised to meet the authenticity constraint:

[0126]

[0127] Update the client participation rate and modify the bid: the final client participation rate update formula is obtained as:

[0128] a i ←a i +Δa i

[0129] b i ←c i +k i ·(Δa i ) 2

[0130] This mechanism ensures that the client evolves reasonable behavior on a self-interest basis, and the system as a whole can maintain training stability and participation persistence without frequent intervention.

[0131] As shown in Figure 2 , the embodiment constructs an actual federated learning system, including a center server and multiple edge clients. The center server is responsible for core operations such as task publishing, model initialization, client evaluation, incentive allocation, and global model aggregation. The edge client is the main body of distributed training, has local data processing capability, and is divided into different categories according to participation preferences. The system describes the client participation behavior through a Markov chain model and sets different participation rates for different types of clients. The client can decide whether to participate in the current round of training according to its own state, and each client has heterogeneous non-identically distributed data labels and different sizes of data, which truly reflects the federated learning scene in a heterogeneous edge environment.

[0132] The embodiment is based on the federated learning system architecture shown in Figure 2 , which is composed of a center server and multiple edge clients. The clients are divided into three categories according to their activity levels and behavior patterns: high activity, low activity, and stable clients. The online state of the client is determined by a Markov chain, and the following steps are performed each round until the global model converges.

[0133] The step process is shown in Figures 1 to 3 , that is, the autonomous adjustment and data balancing federated learning method based on the incentive mechanism provided by the present application, which specifically includes the following steps:

[0134] S1, system initialization and release of training tasks, calculate the label distribution heterogeneity index PCDI: the system reflects the skew degree of the label distribution of the client data, and the PCDI value of the strong heterogeneity client is higher, and the weak heterogeneity client is lower. This step provides the basis for subsequent label distribution-based client screening.

[0135] S2, identify the online client set C of the current training round t : Before each round of training begins, the server detects and identifies the current online client set C t . High availability clients, such as always-on edge servers, almost always appear in C t ; low availability clients, such as battery-dependent and occasionally networked smartphones, only temporarily join when the device is on or the network is good; stable clients, such as smart cameras or home devices that upload data regularly, are online periodically. The server combines real-time network conditions and client responses to dynamically update C t , ensuring that each round of training selects currently available online clients, fitting real distributed environments.

[0136] S3, select a certain proportion of clients in ascending order of PCDI to form the initial candidate set S t : The server sorts the PCDI values of each online client and selects the first ζ proportion of clients with the smallest PCDI to form the initial candidate set S t . This strategy prioritizes weakly heterogeneous clients with more balanced data distribution, ensuring that S t contains the most balanced data distribution among the current online clients.

[0137] S4, based on S t Calculate the initial global label distribution index GCDI: the server aggregates the local data of all clients in set S t , calculate the initial global label distribution consistency index GCDI. Specifically, it is expressed as:

[0138]

[0139] GCDI is used to measure the skewness of the global label distribution formed by the data provided by S t . If S t is mainly composed of weakly heterogeneous clients with balanced label distribution, the GCDI value is low, indicating that the initial global data distribution is relatively balanced; otherwise, if S tIf there is still a certain proportion of label skew data, the GCDI value will be relatively high. The GCDI index comprehensively measures the balance and diversity of global labels, and not only gives priority to clients with balanced label distribution and low heterogeneity, but also values the introduction of clients containing global scarce sample labels, thereby effectively supplementing the lack of global labels. This index ensures that the training set not only maintains low heterogeneity, but also covers rich category information, thereby improving the generalization ability of the global model.

[0140] S5, screening participating clients based on the GCDI growth threshold mechanism: the server uses the initial GCDI calculated in step S4 to determine whether more clients need to be supplemented to participate in training using the growth threshold mechanism. Specifically, after the system introduces clients with large PCDI one by one or in batches, the influence of the newly added clients on the global label distribution is calculated:

[0141]

[0142] If the increment does not exceed the threshold, it means that the newly added clients will not significantly increase the global heterogeneity, and the clients are retained for training; otherwise, if the increment exceeds the threshold, the client is excluded. The present application uses a dynamic threshold mechanism τ = 1 + (1 + 0.01·t)·β. Wherein, t represents the current training round, and β is the adjustment factor. This mechanism sets a lower threshold at the beginning of training, emphasizing the stability of the global model, and avoiding model oscillation caused by excessive heterogeneous data disturbance; as the round increases, the threshold gradually increases, the system gradually relaxes the restriction on the heterogeneity of the client, and more data with diverse labels are introduced, thereby enhancing the generalization ability of the model. This design realizes the smooth transition from "steady-state learning" to "diverse fusion", and improves the adaptability and robustness of the model in a heterogeneous data environment. This mechanism ensures that the training set selected under different online clients in each training round maintains low label heterogeneity and good diversity, and balances data representativeness and training stability.

[0143] S6, complete the construction of the benchmark training client set: after screening and supplementing, the server finally determines the initial benchmark training client set. This set contains not only high-availability clients with balanced label distribution, but also strong-heterogeneity or scarce-label clients introduced to improve data diversity. If some labels are not covered, the server will continue to supplement in subsequent rounds. Overall, this set balances diversity and availability, laying a balanced foundation for subsequent training.

[0144] S7, obtain the client data set size m i and the historical participation frequency f i : After determining the initial candidate set, the server further obtains the local data set size m i and the historical participation frequency f i and other auxiliary indicators of each candidate client. The data set size mi reflects the number of samples owned by the client; historical participation frequency f i , which represents the number of times the client is selected in previous rounds of training, can effectively describe the weight of the client in historical model aggregation. These indicators will be used for subsequent contribution calculation to ensure that the selection of each client considers both data volume and fairness.

[0145] S8, calculate the comprehensive contribution v of the client by integrating multi-dimensional contribution indicators i : the server combines data volume m i , historical participation frequency f i , and label heterogeneity indicators to calculate the comprehensive contribution of each candidate client to measure its potential value in this round of training. Generally, clients with large data volume, high participation frequency, and balanced label distribution have higher contribution; otherwise, clients with small data volume, low participation frequency, or skewed label distribution have relatively low contribution. This comprehensive indicator provides an important basis for subsequent auction screening.

[0146] S9, the client submits the participation bid b i : after the server publishes the participation rules, each candidate client submits the participation bid b i . The bid b i reflects the cost required for the client to participate in training or the expected compensation, which is usually related to factors such as resource consumption, communication overhead, and energy consumption. At the same time, the client will adjust the bid according to its own strategy to achieve the best balance between participation probability and revenue.

[0147] S10, the server sorts the cost-effectiveness and determines the final set through the reverse auction mechanism, and generates the auction boundary value: after receiving all the bids from the candidate clients, the server calculates the cost-effectiveness indicator v i based on the pre-calculated comprehensive contribution v This indicator reflects the ratio of client contribution to cost, which is used to evaluate the cost-effectiveness. The server sorts the rank i from high to low, and selects the clients in reverse auction order until it reaches the resource or participation limit, finally determines the selected client set and the auction boundary value rank s , which is the contribution cost-effectiveness of the last selected client. This indicator can guide the client to adjust the optimal participation strategy according to its own utility function.

[0148] S11, assign aggregation weights p i and training rewards to the selected clients: the server assigns aggregation weights p It depends on multi-dimensional evaluation indicators to ensure the fairness and effectiveness of aggregation. The unselected clients are not assigned weights. In terms of rewards, the server gives rewards according to the bid and ranking, encouraging clients to bid truthfully and continuously participate. Clients with higher contributions usually get greater weights and higher returns, while clients with high bids and low contributions get less rewards. The weights and rewards of different types of clients vary due to their contribution differences: high-availability big data clients have larger weights, strongly heterogeneous clients may get additional incentives for providing new labels, and low-availability clients have relatively low weights and rewards.

[0149] S12, the client performs the current round of training and uploads the model update: all selected clients in the current round perform local training and upload the model parameter update w i t to the server.

[0150] S13, the client obtains the auction boundary value rank s and updates the participation rate based on the utility function: after model aggregation, the server notifies all clients of the current round of auction boundary value rank s . The client calculates the utility function based on the expected return and the calculated energy consumption cost:

[0151]

[0152] to evaluate the net return, and accordingly adaptively adjust the participation probability of the next round, which can be determined by analysis:

[0153]

[0154] According to the above formula, high-availability clients tend to maintain a stable participation rate due to higher adjustment costs; low-availability clients tend to increase their participation rate more due to the potential for marginal revenue; and strongly heterogeneous clients may adjust their strategies to improve their selection probability if the return is low. This mechanism guides clients to rationally optimize their participation behavior, improving system resource utilization efficiency and training stability.

[0155] S14, apply the authenticity constraint to determine the optimal participation rate adjustment and modify the bid according to the new cost: after evaluating the utility and adjusting the participation rate, the client needs to follow the authenticity incentive mechanism to update the optimal participation strategy and bid, i.e. The authenticity constraint requires clients to reflect their actual costs and benefits, avoiding false bids to gain undue advantage. Then the client determines the optimal bid b s after adjusting the participation rate based on the current round boundary value rank i and the cost evaluation result. This mechanism encourages clients to bid honestly in the long run, improving system selection efficiency and overall stability.

[0156] S15, the server aggregates the global model: after receiving the local model updates of all selected clients, the server aggregates the global model according to the aggregation weights p of each client i The weighted average is performed to generate a new round of global model. The aggregation formula is as follows:

[0157]

[0158] wherein, represents the local model uploaded by the client i in the tth round, S t is the set of selected clients in this round.

[0159] S16, enter the next round of iteration, repeat steps S2 to S15, until the global model converges: after completing a round of training and benefit distribution, the system enters the next round of iteration, and re-identifies online clients from step S2, and repeats the process of steps S2 to S15. In each round of iteration, the system preferentially selects high availability and weak heterogeneity clients to ensure training stability; strong heterogeneity clients are included at appropriate times according to the global label diversity requirements to improve model generalization ability. Low availability clients dynamically adjust their participation rate according to their own cost and benefit. With the iteration, the participation strategy and bid of the client are continuously optimized, and the system gradually integrates various types of data distribution information, balances performance and fairness. Finally, when the global model reaches the preset convergence condition, the training process is terminated, and the system obtains a global model that integrates the contributions of various clients.

[0160] The embodiment embodies that while ensuring the performance and fairness of the global model, the data heterogeneity and participation rate difference in the federated training process are effectively regulated through the dynamic client screening and incentive strategy based on the auction mechanism.

[0161] In summary, the present application proposes a federated learning incentive system and method for heterogeneous data and client availability fluctuations, which introduces an incentive strategy based on an auction mechanism, jointly models the participation frequency, data distribution characteristics and computing power of clients, effectively screens and incentivizes high-quality participants, and improves the performance and stability of the global model.

[0162] The above only details the preferred embodiments and principles of the present application, and for ordinary skilled persons in the art, the specific implementation manner can be changed according to the idea provided by the present application, and these changes should be regarded as the protection scope of the present application.

Claims

1. A self-regulating and data-balanced federated learning system based on an incentive mechanism, characterized in that: include: The client-side filtering module with global heterogeneity awareness is used to dynamically evaluate and filter client-side data quality through a global label distribution offset awareness mechanism. The server-side utility evaluation and reverse auction module is used to quantify the actual contribution value of the client to the global model training and to allocate incentive resources through an auction mechanism. The client participation rate adaptive adjustment module is used to guide clients to autonomously optimize their participation frequency based on their own historical benefits and costs, thereby achieving long-term participation behavior, based on utility functions.

2. The incentive-based self-regulating and data-balanced federated learning system according to claim 1, characterized in that, The client-side filtering module for global heterogeneity awareness specifically includes the following processes: Calculate Individual Label Distribution Shift: For each client, calculate the degree of imbalance in the corresponding local label distribution, defined as Individual Label Distribution Shift (PCDI), which measures the degree of deviation of the client's local data from the ideal IID distribution. Where λ(i,l) represents the proportion of label l in client i. For an ideal uniform distribution; Collect and sort the set of online clients: Before the start of each training round, the system collects the set of currently available online clients. and set C t Client-side heterogeneity index by label PCDI i Sort by size from smallest to largest, where C is the complete set of the client collection; Construct the initial client set: prioritize selecting the top ζ·|C t The initial participating set S consists of clients with relatively low data heterogeneity. t ζ represents the proportional parameter that constitutes the client of the low heterogeneity benchmark; Calculate the initial global label distribution index: for the current initial participation set S t Calculate the corresponding weighted global label distribution center, i.e., the global label balance index GCDI, to measure the overall heterogeneity level of the current participating clients: Where, m j Let j be the size of the dataset for client j; Used to correct the bias in the model's aggregated weights; λ j The local dataset of client j is normalized and distributed to evaluate the label offset of the overall dataset formed by the current client set. Add the remaining clients to the set and check the constraints: The server attempts to add the remaining clients to S one by one in sorted order. t And recalculate the new GCDI, denoted as GCDI. S ′ Determine whether the relative growth of the global heterogeneity index after a new client joins satisfies the following conditions: τ=1+(1+0.01·t)·β Where t represents the current training round, and β is the adjustment factor; Remove clients that do not meet the criteria: If the global heterogeneity growth does not exceed the threshold τ, then accept the corresponding client to join the new set S. t Otherwise, the corresponding client will be excluded from the current round of filtering.

3. The incentive-based self-regulating and data-balanced federated learning system according to claim 2, characterized in that, The server-side utility evaluation and reverse auction module specifically includes the following processes: Resume client evaluation metrics: The amount of data (m) collected by the server from each client. i Historical participation frequency f i Label Heterogeneity Index PCDI i η is the weighting balancing factor; Calculate the overall contribution index: Establish a multi-dimensional index to evaluate the overall contribution of each client (v). i : Client submits a quote: Clients interested in participating submit their corresponding quotes to the server. i ; Server-side cost-effectiveness calculation and ranking: The server calculates the cost-effectiveness ranking based on client contribution metrics and pricing. i : According to rank i Sort the clients from highest to lowest, select a number of clients for training, and denote the last client that satisfies the incentive condition under the limited budget B as the auction boundary value, denoted as rank. s ; Assigning aggregate weights and incentives: The server calculates the aggregate weight p for each client. i : The final incentive amount will be determined based on the cost-effectiveness ranking and auction boundary value. Where, r i This represents the reward allocated by the server to client i.

4. The incentive-based self-regulating and data-balanced federated learning system according to claim 3, characterized in that, The client engagement rate adaptive adjustment module specifically includes the following processes: Constructing the client utility function: After each training round, the client constructs the expected utility function u based on the training gains and resource consumption. i as follows: Among them, a i Δa represents the actual participation rate of the client. i Update the client's participation rate by a certain margin; c i A fixed participation cost for each training round, including communication and training costs; k i This indicates the additional cost required to adjust the participation rate; Solving for the optimal participation rate increment: The client aims to maximize its own utility and solves for the optimal participation rate adjustment Δa. i *, thus obtaining the analytical solution: Calculate the authenticity constraint value: Introduce an authenticity constraint to ensure that the cost-effectiveness after client adjustment is not lower than the original value. The specific formula is as follows: Where, Δv i For the updated server to be aware of client utility, Δb i Bidding for updated client applications; Changes in participation rate were adjusted based on constraints: This indicates the maximum allowable adjustment value for the participation rate while maintaining the authenticity of the incentives; if Then the response should be Δa i * Perform cropping and adjustments to meet realism constraints: Update client participation rate and modify bids: The final formula for obtaining the updated client participation rate is: a i ←a i +Δa i b i ←c i +k i ·(Δa i ) 2 。 5. A self-regulating and data-balanced federated learning method based on an incentive mechanism, applied to the self-regulating and data-balanced federated learning system based on an incentive mechanism as described in any one of claims 1-4, characterized in that, The incentive-based self-regulation and data-balanced federated learning method includes the following steps: S1, Initialize the system, issue training tasks, and calculate the client label distribution heterogeneity index PCDI. i ; S2, before the start of each training round, identifies the current set of online clients C. t ; S3, select the top ζ proportion clients in ascending order of PCDI to form the initial participant set S. t ; S4, based on S t Calculate the initial global label balance index (GCDI); S5 uses a GCDI growth threshold mechanism to filter client participation; S6, complete the initial benchmark training client set construction; S7, Get the client dataset size m i Historical participation frequency f i index; S8 calculates the client's overall contribution score v by integrating multiple contribution metrics. i ; S9, Client submits this round's bid. i ; S10, server computing performance The final selection set is determined through a reverse auction mechanism, and the auction boundary value rank is generated. s ; S11, Assign aggregate weight p to the selected client. i and training rewards i ; S12, the client executes this round of training and uploads the model update; S13, The client obtains the current auction boundary value rank. s And based on historical benefits and costs, calculate the utility function and update the participation rate; S14, Apply the authenticity constraint to determine the optimal participation rate adjustment amount and modify the quotation according to the new cost; S15, server-wide aggregation model; S16, proceed to the next iteration, repeat steps S2 to S15 until the global model converges.

Images