This paper introduces ByzEval, a novel framework for Byzantine-robust evaluation functions that enable adaptive consensus thresholds in PoFL systems. It addresses the vulnerability of FL to Byzantine attacks and the limitations of static consensus thresholds. The framework includes multi-dimensional reputation scoring, adaptive threshold mechanisms, and attack-resilient evaluation functions.
Key findings
ByzEval achieves 15–30% higher test accuracy than state-of-the-art Byzantine-robust aggregators under high malicious client ratios.
The adaptive consensus mechanism reduces false positive rates by 60% compared to static thresholds.
Theoretical analysis establishes convergence guarantees under Byzantine settings and derives bounds on the maximum tolerable fraction of malicious clients.
Limitations & open questions
The effectiveness of ByzEval under varying network conditions and attack patterns requires further exploration.
The framework's scalability in extremely large-scale federated learning environments remains to be tested.