Briefing
Verifiable Secret Sharing (VSS) schemes in Distributed Privacy-preserving Machine Learning (DPML) face significant challenges concerning commitment consistency and high computational and communication burdens. This research addresses these issues by first identifying a novel Adaptive Share Delay Provision (ASDP) attack and its accompanying Customized Model Poisoning Attack (ACuMPA), which exploit vulnerabilities in existing Byzantine Fault Tolerant (BFT)-based VSS systems. The paper then proposes EByFTVeS, an Efficient Byzantine Fault Tolerant-based Verifiable Secret-sharing scheme, which leverages a modified Practical Byzantine Fault Tolerance (PBFT) consensus mechanism to enforce strict consistency and timing constraints on share distribution, thereby effectively countering adaptive model poisoning and enhancing the robustness of DPML architectures.
Context
Before this research, established VSS-based DPML schemes struggled with two foundational problems ∞ ensuring consistency of cryptographic commitments and managing substantial computational and communication overheads. While BFT systems were introduced to guarantee consistency and improve efficiency in VSS-based secure multiparty computation, this paper reveals that these systems remained vulnerable to sophisticated model poisoning attacks. Specifically, malicious participants could strategically delay broadcasting meticulously crafted shares, leading to inconsistent shares and compromised model integrity.
Analysis
The core mechanism of EByFTVeS centers on a modified four-phase consensus algorithm, building upon the principles of Practical Byzantine Fault Tolerance (PBFT). This scheme introduces a “Pre-Propose” stage where participants independently initiate requests, batch them, and send initial proposals to a primary node. EByFTVeS mandates that all shares and their commitments undergo this consensus mechanism, compelling participants to submit and verify their shares before aggregation.
This process fundamentally differs from previous approaches by preventing malicious dealers from strategically delaying the broadcast of customized, poisoned shares, thereby ensuring share consistency and effectively neutralizing the Adaptive Share Delay Provision (ASDP) attack. The scheme’s design enforces that only consensus-approved shares are utilized for model aggregation, safeguarding the integrity of the distributed machine learning model.
Parameters
- Core Concept ∞ Verifiable Secret Sharing (VSS)
- New System/Protocol ∞ EByFTVeS (Efficient Byzantine Fault Tolerant-based Verifiable Secret-sharing)
- Key Attack Identified ∞ Adaptive Share Delay Provision (ASDP) and Customized Model Poisoning Attack (ACuMPA)
- Underlying Consensus ∞ Modified Practical Byzantine Fault Tolerance (PBFT)
- Application Domain ∞ Distributed Privacy-preserving Machine Learning (DPML)
Outlook
This research opens new avenues for enhancing the security and robustness of distributed privacy-preserving machine learning systems. The EByFTVeS scheme offers a foundational building block for future protocols, particularly in federated learning and other privacy-sensitive AI applications where resilience against adaptive adversaries is paramount. Future work will likely involve exploring the integration of EByFTVeS with other cryptographic primitives to further optimize efficiency and expand its applicability across diverse distributed computing environments, ensuring the integrity of shared secrets even under advanced attack vectors.