Enhancing IoT Intrusion Detection Systems Through Horizontal Federated Learning and Optimized WGAN-GP

Abstract

The Internet of Things (IoT) ecosystem is fraught with substantial vulnerabilities, particularly in the realm of cybersecurity attacks. Network Intrusion Detection Systems (NIDS) stand as a pivotal element in mitigating these cybersecurity risks. This paper introduces an innovative approach to fortifying IoT security by effectively addressing the data limitations inherent in AI-based NIDS. We present a data generation model that harnesses Generative Adversarial Networks (GANs). Specifically, the GAN variant we employ is Wasserstein GAN with Gradient Penalty (WGAN-GP), which combines the Wasserstein loss formulation with a gradient norm penalty to stabilize training and improve the quality of generated data. The performance is optimized with Genetic Algorithms, focusing on hyper-parameter selection and federated learning for shared model weights. The model’s training is conducted on four well-established benchmark datasets: UNSW-NB15, IoT-23, CSE-CIC-IDS2018, and MQTT-IoT-IDS2020. We conduct a comprehensive comparative analysis between the generated synthetic data and real-world datasets, rigorously assessing their impact on training Machine Learning (ML) models. The findings underscore the efficacy of our approach, demonstrating a significant improvement in detection accuracy, achieving a 99% accuracy rate when combining the generated data with real datasets. This study highlights the paramount significance of innovative techniques in enhancing the security of IoT systems. Furthermore, it presents a promising avenue for generating high-quality synthetic tabular data, despite its complexity and time-consuming implementation. Such data can be leveraged across a large spectrum of applications, including training ML models, data augmentation, and privacy-preserving data sharing.