SwinIoT: A Hierarchical Transformer-Based Framework for Behavioral Anomaly Detection in IoT-Driven Smart Cities

Abstract

Internet of Things (IoT) devices become more and more important because they are useful in different applications, for example, traffic monitoring, public safety, and environmental management. However, the vastness and diversity of IoT data and the requirement for real-time decision-making create a big challenge for anomaly detection frameworks. In this work, we propose SwinIoT, a new framework with hierarchical and windowed attention mechanisms of Swin Transformer that is especially good for the purpose of behavioral anomaly detection in IoT settings. SwinIoT would solve important problems of class imbalance, noisy data, and heterogeneous devices through the introduction of custom attention models embedding real-time optimizations. The proposed framework was benchmarked on nine datasets such as ARAS, CASAS, WESAD, and UCF Crime compared to the above-mentioned state-of-the-art algorithms like Active Learning-Based Anomaly Detection, Deep Support Vector Data Description (DSVDD), Deep Support Vector Data Description Contractive Autoencoder (DSVDD-CAE), and Federated Principal Component Analysis (FedPCA). It is proven to be better than the above algorithms by attaining up to 96% accuracy, 97% Mean Average Precision (mAP), and excellent Area Under the Receiver Operating Characteristic curve (AUC-ROC) as well as Precision-Recall (PR) metrics, especially in low-resource and unbalanced data scenarios. The results indicate the potential of SwinIoT in scalable and accurate anomaly detection to the development of safer, smarter cities, ensuring reliability and security in critical systems.