On the Hidden Costs of Fairness in Byzantine-Robust Decentralized Federated Learning under Data Heterogeneity

Byzantine-robust aggregation methods secure decentralized federated learning (DFL) against malicious clients, yet their fairness effects remain poorly understood. We present the first systematic study of how such defenses impact group fairness in DFL, evaluating six state-of-The-Art methods across three vision datasets under varied attacks and graph topologies. Our results show all robust methods consistently degrade fairness relative to weighted averaging, even without attacks, with degradation worsening under sophisticated attacks like ALIE and DISSENSUS. However, not all defenses degrade fairness equally. We identify the root cause as a fundamental misalignment between outlier-detection mechanisms and fairness objectives, where defenses treat statistical deviation as adversarial, inadvertently penalizing minority client updates. These findings have critical implications for DFL deployment in fairness-sensitive domains, underscoring the need for fairness-Aware Byzantine defenses that can distinguish between malicious and underrepresented statistical patterns.