A Lightweight CNN for CAN-Bus Intrusion Detection

Automotive electronic control units (ECUs) typically offer less than 32 kB of on-chip flash memory, presenting a significant challenge for deploying deep-learning models for Controller Area Network (CAN bus) intrusion detection systems (IDS). While existing deep learning methods demonstrate high efficiency in threat detection, their complex architectures often render them too resource-intensive for the stringent constraints of automotive ECUs. This paper addresses this critical gap by proposing an efficient methodology for developing an ultra-lightweight CNN-based IDS. Our approach centers on the systematic application and validation of an information-dense RGB tensor encoding for 16-frame CAN windows (adapted from prior work [1]), which effectively captures crucial temporal, identifier, and data-length characteristics. This encoding is coupled with CANET-33K, a novel and compact 32.9k-parameter CNN architecture specifically tailored for these RGB representations, and a one-shot hybrid compression strategy that combines structured pruning with 8-bit post-training quantization (PTQ). The resulting INT8 model achieves a 99.73% overall accuracy and a 0.40% false-negative rate across four common attack types (DoS, Fuzzy, Gear, RPM). Notably, the compressed model is over 420 times smaller than a 7 MB Inception-ResNet baseline and 16 times smaller than recent specialized RGB-CNNs, while maintaining comparable or superior detection performance. These results demonstrate a viable path towards deploying high-performance IDS on resource-constrained automotive hardware.