Towards energy-efficient federated anomaly detection in dynamic environments

Federated anomaly detection offers a promising approach to mitigating false positives in non-IID environments with geographically distributed devices. However, dynamic IoT settings are frequently subject to evolving definitions of 'normal' behavior, a phenomenon known as concept drift. Furthermore, because these applications typically run on resource-constrained edge devices, they require highly efficient online retraining mechanisms to adapt to these drifts without draining system resources. In this work, we propose a federated anomaly detection framework tailored for such distributed, dynamic environments. Our approach integrates concept drift detection with adaptive model retraining and decentralized aggregation across multiple devices. We evaluate the proposed framework on synthetic datasets, demonstrating that explicitly addressing concept drift via online retraining significantly enhances anomaly detection performance. Ultimately, these results highlight the viability of accurate, energy-efficient federated anomaly detection in dynamic IoT networks.