Smart Defense: Harnessing Hybrid Deep Learning Models for Resilient IoT Intrusion Detection

Abstract

Internet of Things (IoT) networks have transformed various industries by enabling seamless connectivity and automation, yet they also pose significant security challenges. Traditional intrusion detection systems (IDS) struggle to protect these complex and diverse networks due to the vast variability in IoT devices, protocols, and communication patterns. This study explores the integration of deep learning (DL) and adversarial techniques to enhance IDS performance for IoT network security. We propose a DL-based IDS framework utilizing hybrid models, including convolutional neural network and long-short term memory (CNN-LSTM), bidirectional LSTM (B-LSTM), and bidirectional GRU (B-GRU). Experiments on the ToN-IoT dataset achieved accuracy levels exceeding 98% in non-adversarial scenarios. Among the models, B-LSTM exhibited outstanding resilience to adversarial attacks, such as FGSM, PGD, and Deep Fool, demonstrating its suitability for real-world IoT network security applications. This study highlights the need for robust IDS models to secure IoT networks effectively and emphasizes the importance of rigorous testing against adversarial threats, even when high accuracy is achieved.