Lightweight Autoencoder-Isolation Forest Anomaly Detection for Green IoT Edge Gateways

The rapid growth of the Internet of Things (IoT) has given rise to highly diverse and interconnected ecosystems that are increasingly susceptible to sophisticated cyber threats. Conventional anomaly detection schemes often prioritize accuracy while overlooking computational efficiency and environmental impact, which limits their deployment in resource-constrained edge environments. This paper presents \textit{EcoDefender}, a sustainable hybrid anomaly detection framework that integrates \textit{Autoencoder(AE)}-based representation learning with \textit{Isolation Forest(IF)} anomaly scoring. Beyond empirical performance, EcoDefender is supported by a theoretical foundation that establishes formal guarantees for its stability, convergence, robustness, and energy-complexity coupling-thereby linking computational behavior to energy efficiency. Furthermore, experiments on realistic IoT traffic confirm these theoretical insights, achieving up to 94\% detection accuracy with an average CPU usage of only 22\%, 27 ms inference latency, and 30\% lower energy consumption compared to AE-only baselines. By embedding sustainability metrics directly into the security evaluation process, this work demonstrates that reliable anomaly detection and environmental responsibility can coexist within next-generation green IoT infrastructures, aligning with the United Nations Sustainable Development Goals (SDG 9: resilient infrastructure, SDG 13: climate action).