Reputation-Based Model Aggregation and Resource Optimization in Wireless Federated Learning Systems

Abstract

Federated learning (FL) has received widespread attention from academia and industry because it overcomes traditional security limitations associated with model training data. However, the FL process is vulnerable to manipulation by locally malicious users, who can alter their local data, thus impacting the accuracy of the model’s training outcomes. Meanwhile, optimizing delay in FL needs to take individual client fairness into consideration. In this paper, we present a reputation-based model aggregation and resource optimization framework to enhance the efficiency and reliability of training in wireless FL systems. Particularly, we investigate a total delay minimization problem while ensuring fairness among clients, which jointly optimizes client scheduling, transmit rate, bandwidth proportion, and CPU frequency. Considering the non-convexity and high complexity of the objective function, we decoupled the optimal variables and designed an efficient algorithm. By doing this, the client scheduling policy is obtained by deep reinforcement learning. Then, the transmit rate allocation and bandwidth proportion are derived through the Lagrangian dual method. Finally, we attain the CPU frequency allocation via the adaptive harmony algorithm. Simulation results reveal that our algorithm can establish delay fairness among clients and balance convergence performance and delay.