Work-in-Progress: CANGen: Practical Synthetic CAN Traces Generation Using Deep Generative Models

Realistic CAN traces enable a wide range of automotive security applications including but not limited to anomaly detection, fingerprinting, simulation and testing. However, nowadays these applications are blocked by the sparsity of high-quality, diverse CAN traces. Building on the recent advances of Deep Generative Models (DGMs), we identify two key challenges when off-the-shelf DGMs are applied to CAN traces: generation and evaluation. We design an end-to-end, open-source framework CANGen which efficiently tackles the generation and evaluation of synthetic CAN traces with various DGMs. Our empirical evaluation shows that CANGen can efficiently generate high-fidelity synthetic CAN traces that meet the requirements of domain-specific properties and downstream tasks.