Hybrid aggregation for federated learning under blockchain framework

Abstract

Federated learning based on local differential privacy and blockchain can effectively mitigate the privacy issues of server and provide strong privacy against multiple kinds of attack. However, the actual privacy of users gradually decreases with the frequency of user updates, and noises from perturbation cause contradictions between privacy and utility. To enhance user privacy while ensuring data utility, we propose a Hybrid Aggregation mechanism based on Shuffling, Subsampling and Shapley value (HASSS) for federated learning under blockchain framework. HASSS includes two procedures, private intra-local domain aggregation and efficient inter-local domain evaluation. During the private aggregation, the local updates of users are selected and randomized to achieve gradient index privacy and gradient privacy, and then are shuffled and subsampled by shufflers to achieve identity privacy and privacy amplification. During the efficient evaluation, local servers that aggregated updates within domains broadcast and receive updates from other local servers, based on which the contribution of each local server is calculated to select nodes for global update. Two comprehensive sets are applied to evaluate the performance of HASSS. Simulations show that our scheme can enhance user privacy while ensuring data utility.