Resilient Decentralized Wireless Federated Learning via Gradient Tracking with AdamW

Wireless Internet-of-Things (IoT) edge networks increasingly demand decentralized learning (DecL) methods that remain reliable under two simultaneous stressors: heterogeneous, non-IID local data and communication-constrained wireless links. Existing decentralized optimization schemes, however, tend to incur substantial communication overhead and suffer degraded performance when transmissions are bounded by strict airtime budgets, fading channels, and packet losses.

This paper introduces **QEF-GT-AdamW**, a communication-efficient and outage-resilient algorithm for DecL over wireless communication (WCom) networks. The method is motivated by the observation that robustness in wireless DecL is not merely a matter of compression or of optimization alone, but of jointly addressing data heterogeneity, training stability, payload size, and unreliable broadcast reception.

The proposed design integrates three complementary mechanisms: (i) gradient tracking to mitigate the effect of non-IID data, (ii) AdamW-based adaptive optimization to improve training stability, and (iii) dual-stream biased quantization with error feedback to reduce communication payloads for both model and tracking exc

Experimental evaluation is conducted on heterogeneous MNIST and CIFAR-10 settings, which represent non-IID data distributions across decentralized nodes. The reported results show that QEF-GT-AdamW consistently improves robustness and convergence performance over representative DecL baselines.

Key empirical findings reported in the abstract include:

- **Improved robustness** relative to representative decentralized learning baselines under heterogeneous data and constrained wireless resources.
- **Improved convergence performance**, attributed to the combination of gradient tracking and AdamW-based adaptive optimization.
- **Favorable accuracy-communication trade-offs** under limited wireless resources, enabled by dual-stream biased quantization with error feedback.

The evaluation spans both simpler image classification (MNIST) and more challenging vision tasks (CIFAR-10), suggesting that the benefits are not tied to a single dataset scale. The abstract does not report specific accuracy percentages, communication bit budgets, or convergence rates; those quantitative details would be found in the full paper's experimental tables and figures.