Delay-Attack-Resistant Byzantine Fault-Tolerant Secret Sharing for Distributed Privacy-Preserving Machine Learning

Verifiable Secret Sharing (VSS) is widely employed in Distributed Privacy-Preserving Machine Learning (DPML) to enable participants to detect invalid secret-shares from malicious participants through cryptographic commitment verification. Nevertheless, most existing VSS schemes can hardly combat Byzantine adversaries in maintaining the global consistency of shares. Although Byzantine Fault Tolerance (BFT) mechanisms have recently been integrated into existing VSS schemes to strengthen global consistency guarantees, this paper identifies a previously overlooked vulnerability: the Adaptive Secret-share Delay Provision (ASDP) strategy. We demonstrate how the ASDP strategy can be exploited to mount a Customized Model Poisoning (CMP) attack against any honest participant. A rigorous theoretical analysis elucidates the operational principles of both the ASDP and the CMP within contemporary VSS schemes featuring BFT protection. To counter this new threat, we propose an Efficient Distributed VSS (EDV) scheme. We formally prove the validity, liveness, global consistency, and privacy guarantees of EDV. Extensive experiments confirm that EDV not only neutralizes ASDP-induced vulnerabilities but also achieves superior computational and communication efficiency compared to state-of-the-art VSS schemes.