FedCache: A Knowledge Cache-driven Federated Learning Architecture for Personalized Edge Intelligence

• DOI: 10.36227/techrxiv.23255420.v4
• Publication Date: 2024-02-11T20:59:00Z
• Authors: Zhiyuan Wu; Sheng Sun; Yuwei Wang; Min Liu; Ke Xu; Wen Wang; Xuefeng Jiang; Bo Gao; Jinda Lu
• Keywords: FOS: Computer and information sciences; Computer Science - Distributed, Parallel, and Cluster Computing; 0202 electrical engineering, electronic engineering, information engineering; Distributed, Parallel, and Cluster Computing (cs.DC); 02 engineering and technology

Edge Intelligence (EI) enables Artificial (AI) applications to run at the edge, where data analysis and decision-making can be performed in real-time close sources. To protect privacy unify silos distributed among end devices EI, Federated Learning (FL) is proposed for collaborative training shared AI models across multiple without compromising security. However, prevailing FL approaches cannot guarantee model generalization adaptation on heterogeneous clients. Recently, Personalized (PFL) has drawn growing awareness as it striking a productive balance between local-specific requirements inherent global-generalized optimization objectives satisfactory performance. most existing PFL methods are based Parameters Interaction-based Architecture (PIA) represented by FedAvg, which causes unaffordable communication burdens due large-scale parameters transmission edge server. In contrast, Logits (LIA) update with logits transfer, gains advantages of lightweight on-device allowance compared PIA. Nevertheless, previous LIA attempt achieve performance either relying unrealistic public datasets or increasing overhead additional information other than logits. tackle this dilemma, we propose knowledge cache-driven architecture, named FedCache, reserves cache server fetching personalized from samples similar hashes each given sample. During phase, ensemble distillation applied constructive transferred server-side cache. Empirical experiments four demonstrate comparable FedCache state-of-art approaches, more two orders magnitude improvements efficiency. Our code DEMO available https://github.com/wuzhiyuan2000/FedCache.