Robust Contribution Measurement for Federated Learning

Federated Learning (FL) has emerged as a pivotal framework for privacy-preserving collaborative model training. However, the absence of fair contribution measurement mechanisms fundamentally undermines system integrity, leaving FL vulnerable to both inequitable participant incentives and malicious client behaviors such as free-riding or adversarial attacks. To address these critical challenges, this paper proposes a novel contribution measurement method based on the Banzhaf Index from cooperative game theory, which offers superior computational efficiency and noise robustness compared to conventional approaches. Our framework systematically quantifies each client’s marginal contribution by analyzing performance variations when incorporated into different client coalitions. Experimental results on the MNIST dataset demonstrate the method’s efficacy in effectively distinguishing between normal clients and those exhibiting malicious behaviors. Furthermore, ablation studies confirm the robustness of our approach across diverse system configurations, thereby providing substantial support for client contribution assessment in FL systems.