Machine Learning-Based Intrusion Detection Systems for Automotive In-Vehicle Networks
TL;DR
This survey provides a comprehensive comparative analysis of IDS approaches based on ML, Deep Learning (DL), and Federated Learning (FL), and analyzed these methods not only by detection performance but also according to crucial operational metrics, such as system complexity and applicability in NRT scenarios.
Abstract
The technological trends in smart and autonomous vehicles are shifting the transportation landscape, playing a pivotal role in the future of mobility. Integrating Artificial Intelligence (AI) into smart mobility platforms enhances management, efficiency, sustainability, and safety by enabling data-driven decision-making and timely action. To achieve this, the real-time exchange of data among various vehicle components is of paramount importance. For In-vehicle communication, the Controller Area Network (CAN) bus is the de facto standard solution. Despite its widespread adoption, it lacks some essential security features, such as authentication and encryption of messages, making autonomous vehicles vulnerable to numerous cyber-attacks. In this regard, Intrusion Detection Systems (IDSs) provide an effective solution for detecting attack attempts while providing critical data for mitigation and auditing. This is especially relevant in the context of AI-powered IDS solutions. Although existing surveys in the literature on Machine Learning (ML)-based IDSs for the CAN bus primarily focus on detection accuracy, they critically overlook parameters essential for real-world deployment, such as model complexity, resource utilization, and Near Real-Time (NRT) feasibility. To address these shortcomings, our survey provides a comprehensive comparative analysis of IDS approaches based on ML, Deep Learning (DL), and Federated Learning (FL). Through a structured benchmarking comparison, we analyze these methods not only by detection performance but also according to crucial operational metrics, such as system complexity and applicability in NRT scenarios. Furthermore, we provide a detailed analysis of FL-based approaches for the CAN bus, specifically addressing the critical challenge of data privacy. Our contribution results in a practical roadmap for deploying IDSs that are feasible, resilient, and efficient within real-time, resource-limited automotive environments, thereby offering unique and valuable insights for future research in this domain.