The Controller Area Network (CAN) protocol, commonly utilized in automotive networks, is significantly susceptible due to its absence of security measures, including encryption. Recent breakthroughs in deep learning have been utilized to improve in-vehicle intrusion detection. Nonetheless, implementing these resource-intensive systems on the constrained computational capacities of in-vehicle devices presents significant challenges. This study presents BKIDS, an innovative lightweight intrusion detection method based on transfer learning, specifically designed for the CAN protocol. The proposed BKIDS utilizes a compact DNN trained offline with transfer learning and is deployed for on-device inference without requiring on-device retraining. We optimize the pipeline with lightweight preprocessing and post-training integer quantization, which reduces model size by 56–71% and latency by 59–67% across our tests while preserving detection accuracy. Unlike prior CAN-IDS works that apply transfer learning or quantization in isolation, BKIDS integrates transfer learning and quantization in a single lightweight framework tailored to CAN traffic.