Adversarial Attack on Automotive Radar Point Cloud Classifiers

Automotive radars are critical in autonomous driving, providing robustness in practical, challenging scenarios. Recently introduced deep learning processing dramatically improved automotive radar performance. However, deep learningbased radar perception is vulnerable to adversarial attacks. This work investigates the vulnerability of the deep learning-based radar target classification to adversarial attacks. First, the radarPointNet (R-PointNet), which incorporates radar-specific features, is introduced for the radar target classification using point clouds aggregated over a few frames. Next, the radar-specific adversarial attack, R - C & W, is introduced against the R-PointNet. The performance of the proposed classification network and the adversarial attack efficiency are evaluated using the point clouds from the recorded radar measurements. It was shown that the adversarial attack with indistinguishable perturbation could dramatically degrade radar classification performance, emphasizing the need to develop defense methods against adversarial attacks.