AI-DRIVEN ANOMALY DETECTION IN IOT NETWORKS USING ADVANCED MACHINE LEARNING TECHNIQUES

Abstract

The rapid expansion of Internet of Things (IoT) ecosystems has significantly increased the attack surface, making real-time anomaly detection a critical requirement for ensuring network resilience and security. This research presents a hybrid AI-driven anomaly detection architecture integrating Graph Neural Networks (GNNs), Transformer encoders, and Autoencoder-based reconstruction learning to capture spatial, temporal, and behavioral dependencies within heterogeneous IoT traffic. The model is further enhanced through federated learning, enabling privacy-preserving distributed training across IoT devices while maintaining strong predictive performance. The proposed system is evaluated using benchmark datasets such as N-BaIoT and Bot-IoT, demonstrating superior accuracy, robustness, and cross-device generalization compared to traditional machine learning and standalone deep learning methods.

Comprehensive experiments were conducted across centralized and federated environments to assess detection performance, scalability, model stability, and resilience to concept drift. The hybrid model consistently achieved F1-scores above 98%, outperforming GNN-only, Transformer-only, and Autoencoder-only baselines. Furthermore, the federated version of the model preserved high detection accuracy (99.1%) even under non-IID data distributions, validating its suitability for privacy-sensitive IoT deployments. The results indicate that the fusion of multiple AI techniques, combined with decentralized training, provides a highly effective solution for next-generation IoT network security.