TrustFed: A Reliable Federated Learning Framework With Malicious-Attack Resistance

As a key technology in 6G research, federated learning (FL) enables collaborative learning in resource-constrained edge networks while ensuring individual data privacy. However, traditional federated learning still has many security issues such as model poisoning attacks. The existing defense methods are not robust enough and will bring a lot of overhead to the system, which restricts the application of federated learning in resource-constrained edge networks. To address this issue, in this paper we propose a lightweight hierarchical audit based FL (HiAudit-FL) framework, with aim to enhance the reliability and security of the learning process with minimal system overhead. Model-audit and parameter-audit are alternatively executed to improve audit accuracy while reducing system overhead. Meanwhile, partial audit method is applied for the model audit stage. As the multi-round audit-waiting clients selections is rendered complex by the unpredictability and variability of environmental factors, we model the clients selection process as a partially observable Markov process and propose a diffusion based deep reinforcement learning algorithm to obtain an optimized client selection strategy to minimize audit overhead while ensuring audit accuracy. Simulation results demonstrate that HiAuditFL can effectively identify and handle potential malicious clients accurately, with small system overhead.