Adaptive federated learning secure aggregation scheme based on threshold homomorphic encryption

Abstract

Aiming at the communication bottleneck problem when the current federated learning security aggregation algorithm was applied in a complex network environment, an adaptive federated learning secure aggregation scheme based on threshold homomorphic encryption was proposed.While protecting gradient privacy, users adaptively compress gradients based on the current available bandwidth, greatly reduced communication overhead for federated users.Furthermore, the new dynamic decryption task distribution algorithm and gradient combination algorithm were designed in the phase of aggregation gradient decryption, which relieved the user’s uplink communication pressure.The experimental results show that the proposed scheme can sharply reduce the amount of communication to 4% compared with the existing federated learning scheme with a trivial model accuracy loss of 1%.

Keywords

• federated learning;secure aggregation;gradient sampling;threshold homomorphic encryption