SRFL: A Swarm-Reputation-Based Autonomic Federated Learning Framework for AIoT

Abstract

Federated learning (FL) has emerged as a leading methodology for facilitating collaborative edge learning (EL) across Artificial Intelligence of Things (AIoT) devices, enabling efficient model training and bolstering privacy protection. Nevertheless, current EL methods that depend on trusted servers engender apprehensions concerning potential data leakage and misuse. Moreover, the untrusted AIoT environment increases security threats in EL collaboration. In addressing these challenges, we introduce an innovative swarm reputation (SR)-based decentralized autonomous organization (DAO) autonomous FL framework, SRFL. Within SRFL, we utilize DAO nodes as autonomous units for processing local services, effectively diminishing the communication overhead attributed to frequent interactions, the SR-based DAO committee oversees the FL process and ensures model consistency. SRFL seamlessly integrates FL with the distributed consensus process and introduces an SR-based consensus mechanism to enhance the collaboration process’s trustworthiness. SR utilizes a hierarchical reward and punishment mechanism, designed to equitably reward honest participants and hammer penalize those undermining the system’s stability. Through extensive experimentation with SRFL, employing different models and datasets, we have substantiated its superior performance in efficiency and robustness.