SmartAFL: Enhancing Asynchronous Federated Learning with Staleness-Aware Aggregation and Smart Contracts

Abstract

Federated learning (FL) has emerged as a promising paradigm for training machine learning models across distributed devices while preserving data privacy, finding applications in edge computing, healthcare, and IoT systems. However, the heterogeneity of client devices and asynchronous updates often lead to significant challenges, including computational inefficiency, model inconsistency, and lack of trust in aggregation processes. Existing FL methods, such as synchronous aggregation or fully decentralized approaches, struggle to balance efficiency and consistency, often relying on centralized servers that introduce single points of failure or fail to fairly account for client contributions in asynchronous settings. To address these issues, we propose a novel dual-layer aggregation framework that integrates local asynchronous micro-aggregation with global synchronous macro-aggregation, leveraging blockchain and IPFS for trust and scalability. Our approach dynamically clusters clients based on gradient similarity, employs staleness-aware asynchronous aggregation within clusters, and enforces secure, transparent inter-cluster synchronization via smart contracts. By storing model updates on IPFS and managing metadata on-chain, our method reduces computational overhead while ensuring fairness and tamper-proof execution. Experimental results demonstrate that this framework achieves superior model performance, convergence stability, and equitable contribution tracking compared to traditional FL systems.