Resilient Decentralized Federated Learning for Wireless IoT Networks

Decentralized learning (DecL) in wireless Internet-of-Things (IoT) edge networks faces significant challenges, including heterogeneous local data and unreliable, communication-constrained wireless links. Existing DecL methods often struggle with high communication overhead and degraded performance under strict airtime budgets, fading channels, and packet losses. This research proposes QEF-GT-AdamW, an algorithm designed to be both communication-efficient and resilient to outages. QEF-GT-AdamW integrates several key techniques: gradient tracking to mitigate non-IID data effects, AdamW-based adaptive optimization for training stability, and dual-stream biased quantization with error feedback to reduce the size of model and tracking exchanges. Furthermore, it employs a local fallback strategy to handle unsuccessful packet receptions, ensuring continued operation. The framework explicitly models wireless communication constraints and provides convergence guarantees, demonstrating improved robustness and accuracy over baselines in heterogeneous MNIST and CIFAR-10 settings.