VCFL: A verifiable and collusion attack resistant privacy preserving framework for cross-silo federated learning

To prevent privacy information leakage through model parameters in federated learning, many works use homomorphic encryption to protect clients’ updates. However, most of them result in significant computation and communication overhead. Even worse, few of them have considered the correctness of the aggregated results and collusion attack between internal curious clients and the server. In this paper, we propose VCFL, an efficient verifiable and collusion attack resistant privacy preserving framework for cross-silo federated learning. Firstly, we design a homomorphic signcryption mechanism to sign and encrypt model parameters in one go. Secondly, we employ the blinding technique to resist collusion attack between clients and the server. Moreover, we leverage the batching approach to further reduce its computation and communication overhead. Finally, we simulate VCFL in FedML on real world datasets and models. Extensive experimental results show that VCFL can guarantee model performance while protecting privacy, and it is more efficient in both computation and communication than similar frameworks.