An Adaptive Deep Learning Framework for Energy-Efficient Anomaly Detection in IoT Sensor Networks

Abstract

IoT sensor networks are increasingly deployed in smart environments, generating continuous streams of data that require intelligent and efficient analysis. This paper proposes an adaptive deep learning framework for anomaly detection in IoT sensor data to identify faults, intrusions, and abnormal patterns in real time. The system utilizes an Autoencoder-based deep neural network that dynamically learns normal behavioral patterns and adapts to evolving data distributions, effectively addressing concept drift. A key contribution of this work is the integration of an energy-aware communication strategy that reduces redundant data transmission, thereby extending the lifetime of resource-constrained IoT devices. Furthermore, a lightweight edge–cloud collaborative architecture is introduced, where preliminary anomaly filtering is performed at the edge to minimize latency and bandwidth usage. The model is evaluated using the UNSW-NB15 and IoT-23 datasets, comprising over 2.5 million network flow records with both normal and anomalous classes. Data preprocessing includes normalization, feature selection, and noise filtering, followed by a 70:15:15 train–validation–test split. Unlike traditional threshold-based methods, the proposed framework achieves higher detection accuracy with reduced false positives and demonstrates strong generalization across multiple IoT domains. Experimental results validate improved detection performance, reduced latency, and enhanced energy efficiency.