Effect of the Dual Attention Suppression Attack on the Performance of Self-Driving Car Models - A Preliminary Study

Compared to human-driven cars, full self-driving cars (SDCs) offer greater safety and convenience. However, the liability shifts from the human driver to the car manufacturer, who has to ensure that the SDCs are robust in withstanding any security attack. The Dual Attention Suppression (DAS) attack is a recent visual adversarial attack proposed in the literature that may significantly degrade the performance of an SDC. In this preliminary work, we simulate the DAS attack on SDCs in the CARLA self-driving simulator framework and study the attack's effect on SDC performance. The SDC performance is measured in terms of success rate and driving score. To the best of our knowledge, this is the first work to perform such a study. We apply a visual static patch at a fixed location on the backside of an ambulance vehicle type to induce the DAS attack. The DAS attack reduces the attention of the “car” class from deep learning models for object classification and object detection. We perform several SDC simulations wherein ambulances with and without modification are spawned randomly as neighboring traffic. Experimental results demonstrate that SDC performance (success rate and driving score) is largely unaffected by the DAS attack. While this initial result is promising, further study with different vehicle types and dynamic visual patch locations is required to generalize the result.