Federated Learning Optimization Towards Fairness

During the big data era,the security and privacy of user data are widely concerned,which will cause the dilemma of data sharing.With the development of new technologies,it is necessary to improve people’s lives by using data.However,considering the situation that multiple parties expect to train a model with data sharing.The privacy,security risks,and cooperative fairness issues need to be aware.With the emergence of federated learning,the risk of privacy leakage has been reduced,and the problem that data centralization on the server side has been solved.However,because multiple participants are both cooperative and competitive in federated learning,existing works of federated learning still face challenges of fairness.On the surface,it is unfair that selfish clients and malicious clients attack the global model but have the same profit as clients with active contributions in federated learning.Furthermore,the gap in dataset and computing power of clients will lead to inequity between contributions and benefits,causing the discrimination of vulnerable clients with the unfair client selection strategy,even causing the Matthew effect.In view of the fairness of federated learning,the specific research contents of this paper are as follows:We built a blockchain economic model to transmit the parameters uploading and downloading into credit-based transactions.So that can avoid the phenomenon of the same profits with different contributions.It also takes advantage of blockchain to monitor the clients and record upload and download transactions in the blockchain.In addition,we add a traceability punishment mechanism to ensure the fairness of the learning.Then we use homomorphic encryption to protect the uploaded parameters and introduce the proxy re-encryption to manage the participation rights.Finally,we proved the rationality and robustness of the scheme by game theory.And the simulation results of real datasets showed that it also has a good performance in fairness,security,and accuracy..Secondly,aiming at the problems of participating fairness and benefit fairness in federal learning,we propose a fair federated learning framework based on clustering.The mean idea of clustering is to make vulnerable clients form clusters and increase their chances to participate in " federation ".The global model parameters are allocated according to the contribution to balance the benefits of different participants in the in-cluster learning process.By defining the long-term fairness constraint,we can ensure the long-term stability of the opportunity for all clients to join the cluster.And consider the cosine similarity of the gradient to get good training results in the clustering.Then the Lyapunov method is introduced to transform the clustering problem into an online queue scheduling problem.Finally,we analyzed the influence of the punishment factor on fairness and accuracy.And the results of the experiments proved the effectiveness of the scheme.Finally,aiming at the fairness issues in the client selection of federated learning,a federated learning framework with efficiency and fairness balances is proposed.We use the multi-armed bandit(MAB)learning to predict the local model training time of the clients.Then group the clients based on predicted local training times to reduce aggregation time and improve efficiency.We build a dual-objective optimization function by quantifying the fairness of participation and available bandwidth.So that we can obtain the Pareto optimal solution,ensuring both efficiency and participation rate.Finally,we presented the performance on effectiveness,accuracy,and fairness of the scheme by theoretical analysis and experiments on real datasets.