Privacy Enhanced Federated Learning via Privacy Masks and Additive Homomorphic Encryption

Federated learning has become an effective method to realize collaborative learning among multiple data centers. However, the problem of privacy leakage has not been perfectly solved. We proposed a scheme based on homomorphic encryption with random privacy masks to cope with the problem of private leakage caused by honest but curious participants. Through combining Paillier encryption scheme with privacy mask, security of the learning process is enhanced with a trivial computation and communication cost. Experimental results show that the decryption time is reduced by more than half after using CRT optimization, without degrading the training precision.