Abstract:

[Objective] This paper aims to address issues such as parameter tampering in federated learning models and the inefficiencies associated with communication delays during data training, which are exacerbated by high server concurrency. [Methods] A decentralized hierarchical data sharing framework for federated learning is proposed, integrating cloud agents and employing re-encryption techniques. This framework is designed to leverage block-chain technology to decentralize server operations and to establish a block-chain network at the edge layer, thereby enabling terminal devices to communicate and transmit data with nearby edge devices, reducing time overhead. Based on this block-chain foundation, a cloud agent pool along with a re-encryption sharing algorithm is incorporated to ensure security while selecting optimal agent nodes, facilitating secure data sharing and trusted access control. [Results] Experimental results demonstrate that the proposed sharing framework significantly reduces communication time and model delay compared to traditional federated learning data-sharing frameworks. [Conclusions] The experimental findings indicate that the framework exhibits outstanding performance in terms of transactions per second (TPS), resistance to collusion, controllability, and scalability, ultimately enhancing both the efficiency and security of data sharing.

Key words: federated learning, blockchain, proxy re-encrypti