Blockchain-Based Adaptive Historical Averaging for Client Dropout Resilience in Federated Learning

Abstract

Federated Learning (FL) has emerged as an innovative paradigm that enables heterogeneous and geographically distributed clients to collaboratively train models in a decentralized and privacy-preserving manner. However, FL systems face numerous challenges in real-world deployments, particularly passive participation caused by malicious attacks, where clients drop out due to attacks. This issue, though not intentionally designed by the system, significantly impacts training stability. In this study, we propose BAHA-FL (Blockchain-based Adaptive Historical Averaging Federated Learning. Our approach integrates adaptive historical averaging with exponential decay weighting to effectively compensate for missing parameters due to client dropouts. Our blockchainbased solution ensures the immutability and traceability of model update records, leveraging Distributed Ledger Technology (DLT) to maintain model integrity.