VerifyVFL: Practical Verifiable Vertical Federated Learning

Federated learning enables decentralized model training without sharing raw data, with vertical federated learning (VFL) particularly useful for non-competing organizations with vertically partitioned data. However, existing VFL protocols often assume honest-but-curious behavior, making them vulnerable to malicious actions. Active parties may manipulate gradients, either to save resources or for malicious purposes, while passive parties can launch direct label inference attacks using sample-level gradients. To address these issues, we introduce VerifyVFL, the first practical verifiable neural network-based VFL scheme. Using homomorphic hashing and bilinear pairing, VerifyVFL generates signatures for local model outputs and provides gradient proofs, allowing data owners to verify received gradients. Random masks further protect sensitive sample-level gradients from passive parties. The scheme ensures low overhead, with constant signature and proof sizes, and requires solving an NP-hard problem to forge proofs. Experimental results confirm VerifyVFL’s practical performance, with constant additional overhead.