With the increasing sophistication and frequency of cyber threats targeting modern vehicles [1], the need for efficient, real-time intrusion detection systems (IDS) has become critical. This study explores the feasibility of deploying an autoencoder-based IDS [2] for CAN BUS networks in resource-constrained automotive environments. Unlike conventional deep learning approaches that rely on computationally intensive models, this research focuses on developing a lightweight, optimized autoencoder architecture capable of detecting anomalies while maintaining computational efficiency.
The study follows a systematic approach, beginning with the design of a baseline autoencoder model, followed by iterative improvements in architecture complexity, parameter tuning, and dataset preprocessing to enhance detection performance. Furthermore, the model is deployed on an embedded system using the TinyML framework, enabling real-time execution on low-power microcontrollers. Key challenges such as model compression, inference latency, and memory footprint optimization are addressed to ensure practical implementation.
The proposed IDS is evaluated based on detection accuracy, computational efficiency, and real-world performance metrics, providing insights into the trade-offs associated with deploying deep learning models in embedded automotive environments.