Federal learning method, device and equipment based on multi-dimensional auction
By adopting a multi-dimensional auction-based method in federated learning, filtering the client and sample number, the overfitting problem caused by improper client selection in federated learning is solved, and the fairness and efficiency of learning is improved.
Patent Information
- Application Number
- CN202411891536.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-20
- Publication Date
- 2025-03-11
AI Technical Summary
During the federated learning process, the client's resource and network environment are different in stability, resulting in different training time, affecting the timeliness of model updates, and may lead to overfitting problems.
The federated learning method based on multi-dimensional auction is adopted. By obtaining the bid information of the client, including unit sample training cost, number of samples, training duration and communication rate, the update age value and reputation value are determined based on the preset multi-dimensional auction mechanism, data point evaluation is carried out, and the number of clients and samples that meet the social welfare increment is selected, and the number of clients and samples participating in model training is determined.
It effectively avoids the overfitting problem caused by repeated selection of the same client, improves the fairness and efficiency of federated learning, and fully taps the potential of client data resources.
Smart Images
Figure CN119599099A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer technology, and in particular to a federated learning method, device and equipment based on multi-dimensional auction. Background Art
[0002] Currently, although Federated Learning (FL) is widely used as an effective solution to protect user privacy and solve the data island problem, clients are usually unwilling to provide data for FL training for free, so designing an effective mechanism to motivate clients becomes a key challenge.
[0003] In existing technical solutions, many have adopted a reverse auction-based design to solve this problem. However, due to the different stability of client resources and client network environments, there are differences in training time, which seriously affects the timeliness of model update collection. In addition, some clients may be continuously selected in different communication rounds, which reduces learning performance and causes overfitting problems. Summary of the invention
[0004] In view of this, the purpose of the present invention is to provide a method, device and apparatus for federated learning based on multi-dimensional auction, which can effectively avoid the overfitting problem caused by repeated selection of the same client for training in different rounds in the current federated learning process, thereby improving the fairness of federated learning, fully tapping the potential of client data resources, and further improving the performance and efficiency of learning. The specific scheme is as follows:
[0005] In a first aspect, the present application provides a federated learning method based on multi-dimensional auction, which is applied to a server side, including:
[0006] After releasing the task information to each client, obtaining the bidding information uploaded by each client; the bidding information includes the corresponding unit sample training cost, the number of samples, the training time of the unit sample and the client communication rate;
[0007] Determine the updated age value and reputation value corresponding to each of the clients based on a preset multi-dimensional auction mechanism;
[0008] Evaluate the data points of each client by updating the age value, reputation value and bidding information corresponding to each client, calculate the current social welfare, screen out the clients and sample quantity that meet the current social welfare increment, and determine the participating clients and sample quantity of the current round of global model training;
[0009] After completing local model training using the sample data of the selected sample quantity, the participating client uploads the model update to the server to aggregate into a global model; the server determines the reward information of the participating client and triggers the update operation of age value and reputation value;
[0010] Repeat the above steps until the target accuracy is reached and federated learning is completed.
[0011] Optionally, after publishing the task information to each client, obtaining the bidding information uploaded by each client includes:
[0012] After publishing the task information to each client, the bidding information uploaded by each client after receiving the task information is obtained; the task information includes corresponding data category information, single-round model training duration, budget information and target accuracy.
[0013] Optionally, the triggering of updating the age value and reputation value update operation includes:
[0014] Determine the reputation positive feedback value and the reputation negative feedback value corresponding to each participating client based on each of the training results;
[0015] Determine the reputation value of each participating client by using the interaction freshness discount factor, the reputation positive feedback value corresponding to each participating client, and the reputation negative feedback value, so as to complete the reputation value calculation operation;
[0016] The updated age value is determined based on the previous round of training selection state information corresponding to each of the clients to complete the updated age value calculation operation.
[0017] Optionally, the step of evaluating the data points of each client by updating the age value, reputation value and bidding information corresponding to each client, calculating the current social welfare, screening out clients and sample quantities that meet the current social welfare increment, and determining the participating clients and sample quantities of the current round of global model training, includes:
[0018] Determine the current update age weight factor of each client by using the update age value corresponding to each client;
[0019] Evaluate the sample data points of each client based on the current updated age weight factor, reputation value, single-round model training duration, and bidding information corresponding to each client, so as to obtain data point evaluation results corresponding to each sample data point;
[0020] Based on the preset social welfare maximization principle, budget information, the bidding information, and the data point evaluation results of each client, the current social welfare calculation is performed to screen out clients and sample quantities that meet the current social welfare increment to participate in the current round of federated learning training.
[0021] Optionally, the server determines the reward information of the participating client, including:
[0022] Obtain the model uploaded by the participating clients after completing the local model training using the sample data of the selected sample quantity, update it to the server side and aggregate it into a global model;
[0023] The server determines the reward information corresponding to the participating client based on the model update of the participating client, the corresponding unit data point remuneration information, and the number of samples participating in the local model training.
[0024] Optionally, the server determines the reward information corresponding to the participating client based on the model update of the participating client, the corresponding unit data point remuneration information, and the number of samples participating in the local model training, including:
[0025] For any participating client, the server determines the cost-effectiveness value of the data point ranked first after the selected data point based on the cost-effectiveness ranking of the client, and determines the unit data point remuneration information using the cost-effectiveness value;
[0026] The reward information is determined based on the model update of the participating client, the unit data point remuneration information, and the number of samples participating in the local model training.
[0027] In a second aspect, the present application provides a federated learning device based on multi-dimensional auction, which is applied to a server side and includes:
[0028] An information acquisition module, used to acquire the bidding information uploaded by each client after publishing the task information to each client; the bidding information includes the corresponding unit sample training cost, number of samples, unit sample training time and client communication rate;
[0029] A value determination module, used to determine the updated age value and reputation value respectively corresponding to each of the clients based on a preset multi-dimensional auction mechanism;
[0030] A client screening module is used to evaluate the data points of each client through the updated age value, reputation value and bidding information corresponding to each client, calculate the current social welfare, screen out the clients and sample quantity that meet the current social welfare increment, so as to determine the participating clients and sample quantity of the current round of global model training;
[0031] The value update module is used for the participating clients to complete the local model training using the sample data of the selected sample quantity, and then upload the model update to the server to aggregate into a global model; the server determines the reward information of the participating clients and triggers the update operation of the age value and reputation value;
[0032] The federated learning completion module is used to repeat the above steps until the target accuracy is reached and the federated learning is completed.
[0033] In a third aspect, the present application provides an electronic device, including:
[0034] Memory, used to store computer programs;
[0035] A processor is used to execute the computer program to implement the steps of the aforementioned multi-dimensional auction-based federated learning method.
[0036] It can be seen that in the present application, after the task information is released to each client, the bidding information uploaded by each client is obtained through the server side; the bidding information includes the corresponding unit sample training cost, the number of samples, the training time of the unit sample and the client communication rate; based on the preset multi-dimensional auction mechanism, the updated age value and the reputation value corresponding to each client are determined; the data points of each client are evaluated through the updated age value, reputation value and the bidding information corresponding to each client, and the current social welfare is calculated, and the clients and the number of samples that meet the current social welfare increment are screened out to determine the participating clients and their sample numbers for the current round of global model training; after the participating clients complete the local model training using the sample data of the screened sample number, they upload the model update to the server side to aggregate into a global model; the server side determines the reward information of the participating clients, and triggers the update operation of the age value and reputation value; the above steps are repeated until the target accuracy is reached and the federated learning is completed. That is, in this application, the updated age value and reputation value corresponding to each client are determined by the server side, and the updated age value, reputation value and bidding information uploaded by the client are used to evaluate the data points, and the participating clients and their sample numbers of the current round of global model training are screened out by calculating the current social welfare. Then, after the participating client completes the training, the reward information of the client is determined based on the uploaded model update, and its updated age value and reputation value are updated. After the update is completed, the participating clients and the number of samples participating in the training are re-screened until the target accuracy is reached and the federated learning is completed. In this way, the overfitting problem caused by repeated selection of the same client for training in different rounds in the current federated learning process can be effectively avoided, thereby improving the fairness of federated learning, fully tapping the potential of client data resources, and thereby improving the performance and efficiency of federated learning. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0038] Figure 1 A flowchart of a multi-dimensional auction-based federated learning method provided for this application;
[0039] Figure 2 A schematic diagram of a federated learning process based on multi-dimensional auction provided for this application;
[0040] Figure 3 A schematic diagram of a federated edge learning framework provided for this application;
[0041] Figure 4 A schematic diagram of an updated age value changing over time provided for this application;
[0042] Figure 5 A social welfare comparison diagram provided for this application;
[0043] Figure 6 A schematic diagram of test accuracy comparison under the MNIST dataset provided in this application;
[0044] Figure 7 A schematic diagram of test accuracy comparison under the CIFAR-10 dataset provided in this application;
[0045] Figure 8 A schematic diagram of the structure of a federated learning device based on multi-dimensional auction provided for this application;
[0046] Fig. 9 A structural diagram of an electronic device provided for this application. DETAILED DESCRIPTION
[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0048] Among the existing technical solutions, many have adopted a reverse auction-based design to solve this problem. However, due to the different stability of different client resources and network environments, there are differences in training time, and this seriously affects the timeliness of model update collection. In addition, some clients may be continuously selected in different communication rounds, which reduces the learning performance and causes overfitting problems. To this end, the present application provides a federated learning solution based on multi-dimensional auctions, which can effectively avoid the overfitting problem caused by repeated selection of the same client for training in different rounds in the current federated learning process, thereby improving the fairness of federated learning, fully tapping the potential of client data resources, and thus improving the performance and efficiency of federated learning.
[0049] See also Figure 1 As shown, the embodiment of the present invention discloses a federated learning method based on multi-dimensional auction, which is applied to the server side and includes:
[0050] Step S11, after releasing the task information to each client, obtaining the bidding information uploaded by each client; the bidding information includes the corresponding unit sample training cost, number of samples, unit sample training time and client communication rate.
[0051] Combination Figure 2 As shown, in this embodiment, the server at the server end first publishes task information to each subordinate client, describes the required data category, and sets a round of training time. Then the interested client submits a bid to the server, that is, the bidding information (i.e., b i , represents the bidding information of the i-th client). The step of publishing task information to each client and obtaining bidding information corresponding to the task information uploaded by each client includes: publishing task information to each client; the task information includes corresponding target data category information; configuring the single-round model training duration to obtain the single-round model training duration (represented by T); obtaining the bidding information uploaded by each client after receiving the task information; the bidding information includes the corresponding unit sample training cost (i.e., c i , representing the unit sample training cost of the i-th client), the number of samples (i.e., d i , represents the number of all data points used by the i-th client for bidding), the training time per unit sample (i.e., η i , representing the training time required for the i-th client unit data point) and the client communication rate (i.e., R i , represents the communication rate of the i-th client).
[0052] Step S12: determining the updated age value and reputation value corresponding to each of the clients based on a preset multi-dimensional auction mechanism.
[0053] In this embodiment, the server will determine the update age value and reputation value corresponding to each of the clients according to the preset multi-dimensional auction mechanism. It can be understood that if the current training is the first round of federated learning, that is, there is no historical training round, the update age value and reputation value of the client are both initial values and can be set to 1.
[0054] Step S13, evaluate the data points of each client by updating the age value, reputation value and bidding information corresponding to each client, calculate the current social welfare, screen out the clients and sample quantity that meet the current social welfare increment, and determine the participating clients and their sample quantity for the current round of global model training.
[0055] In this embodiment, the data point evaluation of each client is performed through the update age value, reputation value and bidding information corresponding to each client, and the current social welfare is calculated, and the clients and sample numbers that meet the current social welfare increment are screened out to determine the participating clients and their sample numbers for the current round of global model training, including: using the update age value corresponding to each client to determine the current update age weight factor of each client; based on the current update age weight factor, reputation value, single-round model training duration and bidding information corresponding to each client, the sample data point evaluation of each client is performed to obtain the data point evaluation results corresponding to each sample data point; based on the preset social welfare maximization principle, budget information, bidding information, and the data point evaluation results of each client, the current social welfare is calculated, and the clients and sample numbers that meet the current social welfare increment are screened out to participate in the current round of federated learning training. Among them, as an indicator describing the importance of update, the calculation formula of the Age of Update (AoU) value is:
[0056] ;
[0057] In the formula, Represents the updated age value of the i-th client in the t-th round of federated learning training. Binary variable represents the selection status of the i-th client in the t-th round of federated learning training, represents the selection status of the i-th client in the t-1 round of training, Indicates that the i-th client is selected for model training in the t-th round; Indicates that the i-th client was not selected for model training in round t. The size of reflects the client data resource development and utilization rate, so as to treat each client fairly and give the client with low data resource development and utilization rate the opportunity to participate in training; The larger the value, the lower the development and utilization rate of the client data resources, which can reflect more information updates. Update age weight factor for current:
[0058] ;
[0059] In the formula, C represents the client set under the server, , i∈C. This weight factor will be used together with the client's data training cost and reputation value to jointly decide whether to select the i-th data point of the client for training in the tth round.
[0060] Regarding data point evaluation, taking the kth data point of the i-th client as an example, calculate its data point cost performance value To measure, the calculation formula is:
[0061] ;
[0062] In the formula, λ i represents the target parameter of the i-th client; x i represents the data volume information of the i-th client; α i represents the local model size of the i-th client, but since the local models of all clients are composed of similar elements, it is reasonable to assume that all clients have the same local model size, denoted by α; represents the reputation value of the i-th client in the t-th round of training. It should be noted that this embodiment calculates The purpose is to find the data point with the maximum possible cost-effectiveness value for training. By looping through all clients, data points that can improve the social welfare value are selected to participate in the training. During the iteration process, the server continuously searches and updates the current maximum value, and then records the data points associated with this maximum value and its corresponding client j. This comprehensive plan is used as the basis for calculating social welfare. If the social welfare obtained after adding new data points after the iteration increases, continue to the next iteration. If it becomes smaller, it means that the newly added data points cannot further improve the social welfare, and client j will exit the next iteration. Finally, the maximum social welfare value and client solution are obtained on the path.
[0063] For more information about the selection of specific clients and their sample quantities, see Figure 3 The federated learning architecture shown in the figure often includes a server and a series of client devices (i.e., the clients) The server is responsible for selecting the data points that participate in the training process and integrating the model updates submitted by these clients to build a unified global model F(ω). Assume that the client i∈C has x for model training. i data points, conditional on 0 <x i ≤d i After being selected to participate in the global model training, the client can download the current global model ω from the server and perform the training task using the selected data points. The preset social welfare maximization principle can be expressed as:
[0064] ;
[0065] In the formula, is the revenue function of the server, which is a function of the actual number of training data points x i An increasing concave function; Q represents question; is defined as a binary random variable. If the i-th client successfully uploads the updated model parameters to the server within time T, then ;otherwise, . It is expressed as:
[0066] ;
[0067] In the formula, is the random variable of the i-th client, which is set to follow the parameter In order to maximize social welfare, this embodiment imposes a constraint on the total budget B and transforms (Q1) into (Q2) as follows:
[0068] ;
[0069] Due to the increasing nature of the reward function g(), this problem can be expressed in the following equivalent way:
[0070] .
[0071] Step S14: After completing local model training using sample data of the selected sample quantity, the participating client uploads the model update to the server to aggregate into a global model; the server determines the reward information of the participating client and triggers the update operation of the age value and reputation value.
[0072] In this embodiment, the selected participating client can download the initial global model from the server, and then use the sample data of the sample number corresponding to itself to perform model training, and upload the model update to the server for aggregation. After the server performs the global model update, the server determines the reward information of the participating client, including: obtaining the model update uploaded by the participating client after completing the local model training using the sample data of the sample number screened by the participating client and aggregating it to the server to form a global model; the server determines the reward information corresponding to the participating client based on the model update of the participating client, the corresponding unit data point remuneration information, and the number of samples participating in the local model training.
[0073] Furthermore, the server determines the reward information corresponding to the participating client based on the model update of the participating client, the corresponding unit data point remuneration information, and the number of samples participating in the local model training, including: for any participating client, the server determines the cost-effectiveness value of the data point ranked first after the selected data point based on the client cost-effectiveness ranking, and uses the cost-effectiveness value to determine the unit data point remuneration information; determines the reward information based on the model update of the participating client, the unit data point remuneration information, and the number of samples participating in the local model training.
[0074] Specifically, in a specific implementation, if the i-th client is a participating client and its sample size is h, then its reward calculation formula is:
[0075] ;
[0076] In the formula, It represents the unit data point reward information of the i-th client, and its calculation formula is as follows:
[0077] ;
[0078] In the formula, It refers to the value of the first data point after the hth selected data point of the i-th client in the client cost-effectiveness ranking, where j represents the j-th client and g represents the g-th data point of the j-th client.
[0079] At the same time, regarding the update of age value and reputation value, the triggering of the update of age value and reputation value update operation includes: determining the reputation positive feedback value and reputation negative feedback value corresponding to each participating client based on each training result; using the interaction freshness discount factor, the reputation positive feedback value and the reputation negative feedback value corresponding to each participating client to determine the reputation value of each participating client to complete the reputation value calculation operation; determining the updated age value based on the previous round of training selection state information corresponding to each client to complete the updated age value calculation operation. Among them, see Figure 4As shown, it can be seen that the content described is that the client is selected to participate in the global model training at timestamps t1, t2 and t3.
[0080] Taking the i-th client as an example, the calculation formula for its reputation is:
[0081] ;
[0082] In the formula, the binary<a,b> The reputation feedback value of the server is represented by a binary variable a and b, where the binary variables a and b represent the positive reputation feedback value and negative reputation feedback value of the server to the i-th client in the t-th round, respectively. If the local model update of the client is improved compared with the global model of the previous round, the binary is <1,0>; otherwise, it is <0,1>. When the i-th client participates in the t-th round of training, the feedback value becomes and , k represents the interaction freshness discount factor, which contains an important idea. Because the client with high reputation in the early stage does not mean that it is always equivalent, this embodiment pays more attention to recent reputation feedback than earlier reputation feedback. Among them, r represents the current training round number of federated learning (different from t, t represents the total training round number of federated learning).
[0083] Step S15: Repeat the above steps until the target accuracy is reached and the federated learning is completed.
[0084] In this embodiment, after the reward information is distributed to the participating clients and the age value and reputation value are updated, if the learning goal has not been achieved, a new round of global model training will be started, and the participating clients of this round and their sample quantity will be selected, that is, the process will jump back to the step of screening by updating the age value, reputation value and bidding information corresponding to each of the clients until the federated learning is completed.
[0085] In summary, this embodiment provides an uncertainty-aware auction mechanism for federated learning based on update age (UAMTRA, i.e., an Uncertainty-aware Auction Mechanism that considers Training-time, Reputation, and Age of update, which can also be called a fairness-aware auction mechanism for federated learning based on update age), introduces the update age value as an indicator to measure the time interval required for the server to receive the latest update from the client, introduces the reputation value to objectively reflect the data quality and training status of the client, and formulates an optimization problem, the goal of which is to maximize the expected social welfare by selecting the client and determining the amount of training data for each participating client. In this way, each client can be treated as fairly as possible to avoid the problem of overfitting caused by continuously selecting a few client data points, and the greedy algorithm introduced in batch processing is used to select client data points to ensure that the optimal social welfare is achieved with a smaller time complexity, and at the same time, the unpredictability of the client's training time, data quality differences, and data heterogeneity can be effectively dealt with.
[0086] It can be seen that in this application, after publishing task information to each client, the server side obtains the bidding information uploaded by each client; the bidding information includes the corresponding unit sample training cost, number of samples, training time of unit sample and client communication rate; based on the preset multi-dimensional auction mechanism, the updated age value and reputation value corresponding to each client are determined; the data points of each client are evaluated through the updated age value, reputation value and bidding information corresponding to each client, and the current social welfare is calculated, and the clients and sample numbers that meet the current social welfare increment are screened out to determine the participating clients and their sample numbers for the current round of global model training; after the participating clients complete the local model training using the sample data of the screened sample number, they upload the model update to the server side to aggregate into a global model; the server side determines the reward information of the participating clients, and triggers the update operation of updating the age value and reputation value; the above steps are repeated until the target accuracy is reached and the federated learning is completed. That is, in this application, the updated age value and reputation value corresponding to each client are determined by the server side, and the updated age value, reputation value and bidding information uploaded by the client are used to evaluate the data points, and the participating clients and their sample numbers of the current round of global model training are screened out by calculating the current social welfare. Then, after the participating client completes the training, the reward information of the client is determined based on the uploaded model update, and its updated age value and reputation value are updated. After the update is completed, the participating clients and the number of samples participating in the training are re-screened until the target accuracy is reached and the federated learning is completed. In this way, the overfitting problem caused by repeated selection of the same client for training in different rounds in the current federated learning process can be effectively avoided, thereby improving the fairness of federated learning, fully tapping the potential of client data resources, and thereby improving the performance and efficiency of federated learning.
[0087] Combine the following Figure 5-7 The schematic diagram disclosed in the figure specifically illustrates the technical solution of the embodiment of the present application.
[0088] Regarding the entire federated learning process, different rational mechanisms will have an impact on the convergence speed and test accuracy of model training.
[0089] like Figure 5 , Figure 6 and Figure 7As shown, there are four rational mechanisms, namely, the VCG mechanism (Vickrey-Clark-Groves mechanisms) that only considers the uncertainty of client training time, the greedy Greedy Agl (Greedy Algorithm, also known as the greedy algorithm) mechanism that only considers the uncertainty of client training time, the MRATR (a Multi-dimensional ReverseAuction that considers Training time and Reputation) mechanism that considers both the uncertainty of client training time and reputation, and the UAMTRA mechanism provided in this embodiment that considers the uncertainty of client training time, reputation and update age (AoU) at the same time.
[0090] Combination Figure 5 As shown, it can be seen that when the number of clients changes, the social welfare differences shown by the four mechanisms are not significant. Given that the VCG mechanism aims to maximize social welfare, it can be observed that the UAMTRA mechanism proposed in this embodiment can approach the VCG mechanism in effect, that is, it is close to maximizing social welfare. This means that the UAMTRA mechanism not only meets the requirements of computational efficiency, individual rationality, incentive compatibility and authenticity, but also in most cases, the UAMTRA mechanism can be used as an efficient alternative to the VCG mechanism, and its main advantage is that it can significantly reduce computational complexity.
[0091] Combination Figure 6As shown, it can be seen that by comparing the test accuracy performance of the four auction mechanisms in the MNIST (Mixed National Institute of Standards and Technology database) data set. From the results, it can be seen that the UAMTRA mechanism performs well on the MNIST data set compared with the VCG solution, Greedy Agl and MRATR. Among them, the test accuracy of UAMTRA is improved by 16.1%, 20.8% and 4.4% compared with the VCG solution, Greedy Agl and MRATR, respectively. In addition, the number of communication rounds required for model convergence is reduced by 35.2%, 36.8% and 6.8%, respectively. This is because the UAMTRA mechanism takes into account the uncertainty of client training time, client reputation and AoU. After considering the reputation, as many high-quality data points as possible can be selected for training during the client data point selection process, and the network status of high-reputation clients will also be relatively stable. Therefore, the MRATR mechanism and the UAMTRA mechanism of this embodiment will have better test accuracy and convergence speed than the other two mechanisms. Since AoU is taken into account in the UAMTRA mechanism, it can treat each client as fairly as possible and thus more fully tap the potential of client data resources. Compared with the MRATR mechanism, the performance is improved to a certain extent.
[0092] Combination Figure 7 As shown in the figure, it can be seen that by comparing the test accuracy performance of the four auction mechanisms in the CIFAR-10 (Canadian Institute for Advanced Research-10, a public dataset for image classification) dataset. From the results, it can be seen that the UAMTRA mechanism has a more significant advantage on the CIFAR-10 dataset compared with the VCG solution, Greedy Agl and MRATR. Among them, the test accuracy of UAMTRA is improved by 26.1%, 30.5% and 8.6% compared with the VCG solution, Greedy Agl and MRATR, respectively. In addition, the number of communication rounds required for model convergence is reduced by 38.4%, 40.8% and 7.4%, respectively. From the experimental results, it can be seen that the UAMTRA mechanism still has advantages on larger and more complex datasets. Especially in terms of test accuracy, taking AoU into account, compared with the MRATR mechanism, it can fully tap the potential of client data resources, so that its advantage gradually increases with the increase of training cycles and then tends to be flat.
[0093] See also Figure 8 As shown, the embodiment of the present application also discloses a federated learning device based on multi-dimensional auction, which is applied to the server side and includes:
[0094] The information acquisition module 11 is used to acquire the bidding information uploaded by each client after releasing the task information to each client; the bidding information includes the corresponding unit sample training cost, number of samples, unit sample training time and client communication rate;
[0095] A value determination module 12, configured to determine an updated age value and a reputation value respectively corresponding to each of the clients based on a preset multi-dimensional auction mechanism;
[0096] The client screening module 13 is used to evaluate the data points of each client through the updated age value, reputation value and bidding information corresponding to each client, calculate the current social welfare, screen out the clients and sample quantity that meet the current social welfare increment, so as to determine the participating clients and sample quantity of the current round of global model training;
[0097] The value updating module 14 is used for the participating client to upload the model update to the server to aggregate into a global model after completing the local model training using the sample data of the sample quantity screened by the participating client; the server determines the reward information of the participating client and triggers the update operation of the age value and reputation value;
[0098] The federated learning completion module 15 is used to repeat the above steps until the target accuracy is reached and the federated learning is completed.
[0099] Among them, for more specific working processes of the above-mentioned modules, please refer to the corresponding contents disclosed in the aforementioned embodiments, which will not be repeated here.
[0100] It can be seen that in this application, the updated age value and reputation value corresponding to each client are determined by the server side, and the updated age value, reputation value and bidding information uploaded by the client are used to evaluate the data points, and the participating clients and their sample numbers of the current round of global model training are screened out by calculating the current social welfare. Then, after the participating client completes the training, the reward information of the client is determined based on the uploaded model update, and its updated age value and reputation value are updated. After the update is completed, the participating clients and the number of samples participating in the training are re-screened until the target accuracy is reached, and federated learning is achieved. In this way, the overfitting problem caused by repeated selection of the same client for training in different rounds in the current federated learning process can be effectively avoided, thereby improving the fairness of federated learning, fully tapping the potential of client data resources, and thereby improving the performance and efficiency of federated learning.
[0101] In some specific embodiments, the information acquisition module 11 may specifically include:
[0102] The bidding information acquisition unit is used to obtain the bidding information uploaded by each client after receiving the task information after publishing the task information to each client, and configure the single-round model training duration to obtain the single-round model training duration; the task information includes corresponding target data category information.
[0103] In some specific embodiments, the value updating module 14 may specifically include:
[0104] A reputation value acquisition unit, used to determine a reputation positive feedback value and a reputation negative feedback value corresponding to each participating client based on each of the training results;
[0105] A reputation value updating unit, used to determine the reputation value of each participating client by using the interaction freshness discount factor, the reputation positive feedback value and the reputation negative feedback value corresponding to each participating client, so as to complete the reputation value calculation operation;
[0106] The age value updating unit is used to determine the updated age value based on the previous round of training selection state information corresponding to each of the clients to complete the updated age value calculation operation.
[0107] In some specific embodiments, the client screening module 13 may specifically include:
[0108] A weight factor determination unit, configured to determine a current update age weight factor of each of the clients using an update age value corresponding to each of the clients;
[0109] A data point evaluation unit, configured to evaluate the sample data points of each client based on the current update age weight factor, the reputation value, the single-round model training duration, and the bidding information corresponding to each client, so as to obtain data point evaluation results respectively corresponding to each sample data point;
[0110] A screening unit is used to perform current social welfare calculation based on a preset social welfare maximization principle, budget information, the bidding information, and the data point evaluation results of each of the clients, and screen out clients and sample quantities that meet the current social welfare increment to participate in the current round of federated learning training.
[0111] In some specific embodiments, the value updating module 14 may specifically include:
[0112] A model aggregation unit is used to obtain the model uploaded by the participating clients after completing the local model training using the sample data of the number of samples screened by them, and update it to the server side to aggregate it into a global model;
[0113] The reward determination unit is used for determining the reward information corresponding to the participating client on the server side based on the model update of the participating client, the corresponding unit data point remuneration information, and the number of samples participating in the local model training.
[0114] In some specific embodiments, the reward determination unit may specifically include:
[0115] The unit data point remuneration information acquisition subunit is used for determining the cost-effectiveness value of the first data point after the selected data point based on the cost-effectiveness ranking of the client for any participating client, and determining the unit data point remuneration information by using the cost-effectiveness value;
[0116] The reward determination subunit is used to determine the reward information based on the model update of the participating client, the unit data point remuneration information, and the number of samples participating in the local model training.
[0117] Furthermore, the present application also discloses an electronic device. Fig. 9 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content in the diagram cannot be regarded as any limitation on the scope of use of the present application.
[0118] Fig. 9 A schematic diagram of the structure of an electronic device 20 provided in an embodiment of the present application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 is used to store a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the multi-dimensional auction-based federated learning method disclosed in any of the aforementioned embodiments. In addition, the electronic device 20 in this embodiment may specifically be an electronic computer.
[0119] In this embodiment, the power supply 23 is used to provide working voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and the external device, and the communication protocol it follows is any communication protocol that can be applied to the technical solution of the present application, and is not specifically limited here; the input and output interface 25 is used to obtain external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs and is not specifically limited here.
[0120] In addition, the memory 22, as a carrier for storing resources, can be a read-only memory, a random access memory, a disk or an optical disk, etc. The resources stored thereon can include an operating system 221, a computer program 222, etc., and the storage method can be temporary storage or permanent storage.
[0121] The operating system 221 is used to manage and control the hardware devices and computer program 222 on the electronic device 20, which can be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program that can be used to complete the multi-dimensional auction-based federated learning method performed by the electronic device 20 disclosed in any of the aforementioned embodiments, the computer program 222 can further include a computer program that can be used to complete other specific tasks.
[0122] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.
[0123] The technical solution provided by the present application is introduced in detail above. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea. At the same time, for general technicians in this field, according to the idea of the present application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.
Claims
1. A federated learning method based on multi-dimensional auction, characterized in that: Applied to the server side, including: After releasing the task information to each client, obtaining the bidding information uploaded by each client; the bidding information includes the corresponding unit sample training cost, the number of samples, the training time of the unit sample and the client communication rate; Determine the updated age value and reputation value corresponding to each of the clients based on a preset multi-dimensional auction mechanism; Evaluate the data points of each client by updating the age value, reputation value and bidding information corresponding to each client, calculate the current social welfare, screen out the clients and sample quantity that meet the current social welfare increment, and determine the participating clients and sample quantity of the current round of global model training; After completing local model training using the sample data of the selected sample quantity, the participating client uploads the model update to the server to aggregate into a global model; the server determines the reward information of the participating client and triggers the update operation of age value and reputation value; Repeat the above steps until the target accuracy is reached and federated learning is completed.
2. The multi-dimensional auction-based federated learning method according to claim 1, characterized in that: After releasing the task information to each client, obtaining the bidding information uploaded by each client includes: After publishing the task information to each client, the bidding information uploaded by each client after receiving the task information is obtained; the task information includes corresponding data category information, single-round model training duration, budget information and target accuracy.
3. The multi-dimensional auction-based federated learning method according to claim 1, characterized in that: The triggering of updating the age value and reputation value update operation includes: Determine the reputation positive feedback value and the reputation negative feedback value corresponding to each participating client based on each of the training results; Determine the reputation value of each participating client by using the interaction freshness discount factor, the reputation positive feedback value corresponding to each participating client, and the reputation negative feedback value, so as to complete the reputation value calculation operation; The updated age value is determined based on the previous round of training selection state information corresponding to each of the clients to complete the updated age value calculation operation.
4. The multi-dimensional auction-based federated learning method according to claim 1, characterized in that: The data point evaluation of each client is performed by updating the age value, reputation value and bidding information corresponding to each client, and the current social welfare is calculated, and the clients and sample quantity that meet the current social welfare increment are screened out to determine the participating clients and sample quantity of the current round of global model training, including: Determine the current update age weight factor of each client by using the update age value corresponding to each client; Evaluate the sample data points of each client based on the current updated age weight factor, reputation value, single-round model training duration, and bidding information corresponding to each client, so as to obtain data point evaluation results corresponding to each sample data point; Based on the preset social welfare maximization principle, budget information, the bidding information, and the data point evaluation results of each client, the current social welfare calculation is performed to screen out clients and sample quantities that meet the current social welfare increment to participate in the current round of federated learning training.
5. The multi-dimensional auction-based federated learning method according to any one of claims 1 to 4, characterized in that: The server determines the reward information of the participating client, including: Obtain the model uploaded by the participating clients after completing the local model training using the sample data of the selected sample quantity, update it to the server side and aggregate it into a global model; The server determines the reward information corresponding to the participating client based on the model update of the participating client, the corresponding unit data point remuneration information, and the number of samples participating in the local model training.
6. The multi-dimensional auction-based federated learning method according to claim 5, characterized in that: The server determines the reward information corresponding to the participating client based on the model update of the participating client, the corresponding unit data point remuneration information, and the number of samples participating in the local model training, including: For any participating client, the server determines the cost-effectiveness value of the data point ranked first after the selected data point based on the cost-effectiveness ranking of the client, and determines the unit data point remuneration information using the cost-effectiveness value; The reward information is determined based on the model update of the participating client, the unit data point remuneration information, and the number of samples participating in the local model training.
7. A federated learning device based on multi-dimensional auction, characterized in that: Applied to the server side, including: An information acquisition module, used to acquire the bidding information uploaded by each client after publishing the task information to each client; the bidding information includes the corresponding unit sample training cost, number of samples, unit sample training time and client communication rate; A value determination module, used to determine the updated age value and reputation value respectively corresponding to each of the clients based on a preset multi-dimensional auction mechanism; A client screening module is used to evaluate the data points of each client through the updated age value, reputation value and bidding information corresponding to each client, calculate the current social welfare, screen out the clients and sample quantity that meet the current social welfare increment, so as to determine the participating clients and sample quantity of the current round of global model training; The value update module is used for the participating clients to complete the local model training using the sample data of the selected sample quantity, and then upload the model update to the server to aggregate into a global model; the server determines the reward information of the participating clients and triggers the update operation of the age value and reputation value; The federated learning completion module is used to repeat the above steps until the target accuracy is reached and the federated learning is completed.
8. An electronic device, characterized in that: include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the multi-dimensional auction-based federated learning method as described in any one of claims 1 to 6.