PRIVACY-PRESERVING DECENTRALIZED INTRUSION DETECTION SYSTEM FOR IOT DEVICES USING DEEP LEARNING

Abstract

The convergence of advanced networking, breakthrough distributed systems technologies, and smart services has rapidly expanded the threat landscape for IoT devices. Researchers have been looking into lightweight and adaptive technologies to solve the problems of cybersecurity in dynamic smart IoT systems, as these domains are increasingly targeted by cyber-criminals. In most of the scenarios, a peripheral defense, Intrusion Detection System (IDS) is proved effective to protect IoT devices. However, existing intrusion detection techniques have centralized designs with repetitive pre-processing steps, privacy leaks due to raw data exchange, and computationally expensive workloads for the resource constrained IoT devices. In this dissertation, we propose using Deep Learning (DL) and relevant distributed Artificial Intelligence (AI) techniques to develop an efficient and secure distributed IDS model. First, we demonstrate that effective pre-processing of input data greatly reduces the burden on the classifier and enhances accuracy in incremental distributed learning. The first contribution in this dissertation proposes a novel pre-processing technique, which ensures privacy of data of the IoT devices, eliminates redundancies, and selects unique features by following innovative extraction techniques. Our privacy-preserving incremental AI-based IDS can tackle zero-day attacks, particularly mutations of existing attacks. Second, the data imbalance issues in intrusion detection can degrade the model accuracy, particularly in rare classes. To this end, Generative Adversarial Network (GAN) is effective in data augmentation to balance the available training data. The second contribution in this dissertation models the proposed distributed IDS in an innovative manner using Federated Learning (FL), which minimizes the data sharing to enhance privacy and performance. Our approach "FEDGAN-IDS" uses FL and GAN to effectively detect cyber threats in smart IoT systems. This is achieved by distributing the GAN network over IoT devices for training the model using local data and handling the model's distribution using FL. Overall, this dissertation proposes a privacy-preserving distributed IDS for IoT devices suitable for real-time protection scenarios. We evaluate our work using accuracy, delay, and other critical criteria using multiple datasets, such as NSL-KDD and KDD99. The model performs better and converges faster than the state-of-the-art standalone IDS models.