A Research On Incentive Mechanism Of Federated Learning

The data silo problem is one of the common challenges in machine learning.Some companies have a large amount of data that can be used to train machine learning models,which makes it possible to achieve high-performance models through cooperation between companies.However,due to commercial competition and privacy regulations,companies are usually unwilling to share their large data samples with other companies for training machine learning models,resulting in the data silo problem.To solve this problem,crosssilo federated learning has been proposed as a framework that enables multiple companies to cooperate in training machine learning models without revealing their original data.In cross-silo federated learning,participating companies or organizations use local data to train local models and periodically push model updates to a global parameter server for model aggregation,instead of uploading raw data to participate in the training process.This method can effectively break data silos while protecting data privacy.However,because each company obtains different benefits from the models trained through cross-silo federated learning and incurs different costs in the training process,there is a mismatch between the contributions made by companies during the training process and their ultimate profits.This unfairness makes companies unwilling to participate in cross-silo federated learning and hinders the implementation of related applications.Ensuring the fairness of all participating companies in the training process and making collaborative training of machine learning models attractive to them has become a key issue that needs to be urgently addressed.To address the above problems,this thesis proposes a cross-silo federated learning incentive mechanism based on the Vickrey-ClarkeGroves(VCG)method,called the Social Welfare Efficient Cross-silo FL incentive mechanism.This thesis analyzes its theoretical performance and proves that the incentive mechanism can effectively motivate companies to participate in cross-silo federated learning and achieve overall profit maximization,incentive compatibility,individual rationality,and weak budget balance.In addition,the authors of this thesis conducted a large number of experiments on a real test platform based on the MNIST and CIFAR-100 datasets.The experimental results demonstrate the effectiveness of the proposed mechanism under different data distributions.This thesis also introduces a reputation-based punishment mechanism for the proposed mechanism to improve the fairness of the system and uses blockchain technology to solve the problem of traditional cross-silo federated learning’s dependence on a secure and trustworthy third-party organization.