ASR-Fed: agnostic straggler-resilient semi-asynchronous federated learning technique for secured drone network

Abstract

Federated Learning (FL) has emerged as a transformative artificial intelligence paradigm, facilitating knowledge sharing among distributed edge devices while upholding data privacy. However, dynamic networks and resource-constrained devices such as drones, face challenges like power outages and network contingencies, leading to the straggler effect that impedes the global model performance. To address this, we present ASR-Fed, a novel agnostic straggler-resilient semi-asynchronous FL aggregating algorithm. ASR-Fed incorporates a selection function to dynamically utilize updates from high-performing and active clients, while circumventing contributions from straggling clients during future aggregations. We evaluate the effectiveness of ASR-Fed using two prominent cyber-security datasets, WSN-DS, and Edge-IIoTset, and perform simulations with different deep learning models across formulated unreliable network scenarios. The simulation results demonstrate ASR-Fed's effectiveness in achieving optimal accuracy while significantly reducing communication costs when compared with other FL aggregating protocols.