Vertical Federated Learning (VFL), which draws attention because of its ability to evaluate individuals based on features spread across multiple institutions, encounters numerous privacy and security threats. Existing solutions often suffer from centralized architectures, and exorbitant costs. To mitigate these issues, in this paper, we propose SecureVFL, a decentralized multi-party VFL scheme designed to enhance efficiency and trustworthiness while guaranteeing privacy. SecureVFL uses a permissioned blockchain and introduces a novel consensus algorithm, Proof of Feature Sharing (PoFS), to facilitate decentralized, trustworthy, and high-throughput federated training. SecureVFL introduces a verifiable and lightweight three-party Replicated Secret Sharing (RSS) protocol for feature intersection summation among overlapping users. Furthermore, we propose a (42)-sharing protocol to achieve federated training in a four-party VFL setting. This protocol involves only addition operations and exhibits robustness. SecureVFL not only enables anonymous interactions among participants but also safeguards their real identities, and provides mechanisms to unmask these identities when malicious activities are performed. We illustrate the proposed mechanism through a case study on VFL across four banks. Finally, our theoretical analysis proves the security of SecureVFL. Experiments demonstrated that SecureVFL outperformed existing multi-party VFL privacy-preserving schemes, such as MP-FedXGB, in terms of both overhead and model performance.
