Double verifiable privacy-preserving federated learning under multiple keys

Federated learning allows users to locally train the raw data and pass the trained gradients to cloud server to get global model, protecting the data privacy. Most of the existing privacy preserving federated learning schemes use homomorphic encryption to protect gradient privacy, where all users hold the same private key. However, the gradient privacy of other users will be leaked, once any user’s key is compromised. Meanwhile, these schemes suffer from poor model accuracy and efficiency. Therefore, this paper proposes double verifiable privacy-preserving federated learning under multiple keys, in which Paillier homomorphic encryption is utilized to protect the user local data privacy. Furthermore, by combining verifiable secret sharing and linear homomorphic hash, our scheme enables users to verify the correctness of the random number shares and aggregation result, achieving double verifiability. The security analysis indicates that the proposed scheme meets the necessary security requirements, while the performance analysis demonstrates that our scheme reduces computation and communication costs in comparison with existing schemes.