AI-based Anomaly Detection in Smart City IoT Networks

Abstract

Smart cities rely on vast, heterogeneous IoT deployments—ranging from traffic sensors to environmental monitors—to deliver real-time services. However, the diversity and scale of these networks introduce novel security challenges, including spoofing, distributed denial-of-service, and stealthy data drifts. This paper systematically reviews state-of-the-art machine-learning and deep-learning techniques for anomaly detection in smart-city IoT, categorizing approaches into supervised, unsupervised, and hybrid models. We examine classical algorithms (one-class SVM, k-NN), neural architectures (autoencoders, variational autoencoders), and emerging edge-AI implementations that perform inference on resource-constrained devices. Through analysis of representative case studies, we compare system architectures, datasets, and evaluation metrics such as detection accuracy, false-alarm rate, and computational overhead. We identify critical challenges—data scarcity, model drift, privacy concerns, and deployment constraints—and discuss promising directions, including federated learning, explainable models, and ultra-lightweight neural networks. Our review aims to guide researchers and practitioners in selecting and designing adaptive anomaly detectors that balance detection performance with feasibility in real-world smart-city environments.