zkVFL: Verifiable Federated Learning for Free-Rider Attacks via Efficient Zero-Knowledge Proofs

Abstract

Federated Learning (FL) enables model training on distributed devices while preserving data privacy. However, malicious clients can submit fabricated model updates to fraudulently obtain training rewards, a behavior known as free-rider attacks. Existing detection-based solutions analyze anomalies in model updates but lack direct evidence of local training, making it fail to fully prevent free-riders. To address this limitation, we propose zkVFL, a verifiable FL framework leveraging Zero-Knowledge Proofs (ZKP) to ensure the integrity of local training while preserving privacy. To reduce the computational overhead of proof generation in ZKP, zkVFL introduces two novel techniques: (i) anomaly-aware client sampling to selectively perform ZKP verification and (ii) A recursive ZKP protocol (ReMPoT), incorporating a pruning-based layer selection technique, reduces proof generation costs. Experimental results demonstrate that zkVFL improves the accuracy and convergence of FL training under free-rider attacks while significantly reducing the computational and memory overhead of proof generation on resource-constrained devices.