DEEP LEARNING FOR IOT INTRUSION DETECTION SYSTEMS: ANALYTICAL STUDY

Abstract

The exponential spreading of Internet of Things (IoT) devices has dramatically expanded the attack surface of modern networks, rendering traditional intrusion detection systems (IDS) inadequate against increasingly sophisticated cyber threats. This paper presents a comprehensive comparative study of state-of-the-art deep learning (DL)-based IDS approaches for IoT security, examining Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), hybrid CNN-LSTM architectures, Deep Neural Networks (DNN), Autoencoders (AE), and Federated Learning (FL)-based frameworks. We systematically evaluate these approaches across five widely adopted benchmark datasets: NSL-KDD, UNSW-NB15, BoT-IoT, TON-IoT, and CICIoT2023 and assess their performance in terms of accuracy, precision, recall, F1-score, and computational overhead. Two real-world case studies, one of them is 2016 Mirai Botnet attack on Dyn DNS infrastructure and the 2021 Oldsmar water treatment plant intrusion are employed to illustrate the practical limitations of conventional security approaches and to demonstrate how AI-driven IDS could have mitigated these incidents. Our analysis reveals that hybrid CNN-LSTM models consistently achieve detection rates exceeding 99\% while maintaining low false positive rates, but significant trade-offs exist between detection accuracy and resource efficiency for edge-deployed IoT environments. We further identify critical open challenges including class imbalance, concept drift, adversarial robustness, and dataset obsolescence and propose future research directions centred on lightweight models, federated learning, and transfer learning for real-world IoT deployments.