A Distributed Threshold Additive Homomorphic Encryption for Federated Learning with Dropout Resiliency Based on Lattice

In federated learning, a parameter server needs to aggregate user gradients and a user needs the privacy of their individual gradients. Among all the possible solutions, additive homomorphic encryption is a natural choice. As users may drop out during a federated learning process, and an adversary could corrupt users and the parameter server, a dropout-resilient scheme with distributed key generation is required. We present a lattice based distributed threshold additive homomorphic encryption scheme with provable security that could be used in the federated learning. The evaluation shows that our proposal has a lower communication overhead among all lattice based proposals when the number of users in FL exceeds 26.