Large-Scale Mean-Field Federated Learning for Detection and Defense against Byzantine Attacks in IoT

Contribution: This paper is the first study, based on Mean Field Game (MFG) to detect and defend against Byzantine attacks in Federated Learning.

Abstract: Federated Learning (FL) protects data privacy by sharing gradients across clients rather than local training data. However, malicious clients (e.g., attackers, stragglers) hiding in large-scale FL will severely reduce the learning performance. Thus, how to efficiently detect and defend Byzantine attacks in large-scale FL remains an urgent issue. This paper proposes a reputation-aware Mean-Field-Game-based FL (FedMFG) framework, which aims to defend against Byzantine attacks and improve learning performance. Precisely, we first model the process of large-scale FL as a mean-field game problem across clients and prove the existence and uniqueness of mean-field FL gradients (i.e., Nash equilibrium). We then design a mean-field FL gradient calculation algorithm based on stochastic differential equations, i.e., HJB and FPK equations. Based on comparing the cosine similarity of obtained mean-field and individual FL gradients, we build reputation-aware malicious client detection and defence mechanism, which improves the Byzantine robustness of FL with the global learning performance guarantee. Finally, experimental results show that our proposed framework outperforms the baseline algorithms in realizing Byzantine robustness and improving learning performance. Specifically, our algorithm improves the model accuracy by 10% and 72.7% for the no-attacker and attacker scenarios, respectively.

Recommended citation: Kangkang Sun, Lizheng Liu, Qianqian Pan, Jianhua Li and Ju Wu, "Large-Scale Mean-Field Federated Learning for Detection and Defense: A Byzantine Robustness Approach in IoT," in IEEE Internet of Things Journal, doi: 10.1109/JIOT.2024.3409610. (First Author, IF：8.9)
Download Paper