Enhancing federated learning for IoT-based anomaly detection: A reputation-based client selection approach

Federated Learning (FL) enables collaborative model training across decentralized, privacy-sensitive environments but often suffers from slow convergence, unbalanced client selection, and non‑IID data challenges. We propose a Reputation‑Based Client Selection Mechanism with proportional fairness, computing each client’s reputation from accuracy, consistency, network conditions, data quality, and historical reliability. By adaptively prioritizing high‑contributing clients while ensuring equitable participation, our method accelerates convergence and balances contributions. Evaluations on the UNSW‑NB15 intrusion detection dataset under IID and non‑IID settings demonstrate that our approach significantly reduces the number of communication rounds needed to reach stable accuracy compared to FedAvg and FedProx, while enhancing model generalization and robustness. This scalable strategy advances FL for efficient, inclusive learning in IoT and cybersecurity.