Practical Byzantine-Resilient Federated Learning via Zero-Knowledge Proofs and Adaptive Reputation

Abstract

Trust management under adversarial uncertainty is a central challenge in distributed learning systems. We propose Verifiable FL with Two-Stage Selection (VFL-TS), a knowledge-driven framework for Byzantine-resilient federated learning that maintains a dynamic trust knowledge base—updated through cryptographically verified evidence—to guide adaptive client selection. VFL-TS combines multi-dimensional quality signals for heterogeneity-aware client scoring, trust-aware probabilistic sampling, layered zero-knowledge proofs with selective auditing, and secure multi-party computation aggregation with robust validation-loss filtering. Evaluated across 36 adversarial scenarios on MNIST, Fashion-MNIST, and CIFAR-10, VFL-TS consistently outperforms FLTrust—the strongest noncryptographic baseline—with statistically significant accuracy gains (𝑝 < 0.01) and 2.4× faster convergence on MNIST. Malicious client detection is reliable across attack types and intensities (TPR 0.82–0.94, FPR 0.08–0.15). On CIFAR-10, VFL-TS matches or exceeds FLTrust in 7 of 12 scenarios; a failure mode under extreme gradient-sign attacks at high malicious ratios is fully characterized and motivates future work on reputation warm-starting. Selective auditing—verifying only 20% of clients with full proofs per round—reduces expected cryptographic overhead by ≈70% while maintaining 89% cumulative attack-detection probability over 10 rounds. Communication overhead remains below 0.5% of model-update traffic, and server-side proof verification costs under 4 s per client, demonstrating that strong cryptographic accountability is compatible with practical federated learning deployment.Submitted to Knowledge-Based Systems (Elsevier), under review.