FedKAD Boosts IoT Anomaly Detection with Efficient Federated Learning.

Distributed IoT systems continuously generate vast amounts of multivariate time-series data from various sensors and devices. Detecting anomalies in these streams is crucial for maintaining system health, predicting failures, and ensuring security. However, traditional anomaly detection methods often struggle with the decentralized, non-IID nature of IoT data, as well as the limited computational and communication resources of edge devices. This paper introduces FedKAD, a novel federated Koopman anomaly detection framework designed specifically for these challenges. Unlike resource-intensive deep learning models, FedKAD employs lightweight sliding-window Koopman representations to learn normal temporal dynamics. This approach significantly reduces the computational burden on edge devices. The federated training process in FedKAD is framed as a low-rank consensus problem. Crucially, raw sensor data and local dynamic representations remain on the device, with only compact subspace variables exchanged with a central server. The framework utilizes a federated Stiefel-ADMM algorithm for optimization, ensuring convergence even with partial client participation. During inference, each client performs local anomaly detection by comparing observed trajectories with the learned Koopman dynamics. Experimental results show FedKAD matches or surpasses deep learning baselines in performance while offering substantial gains in training speed, communication efficiency, and inference latency, making it highly suitable for constrained IoT environments.