An Adversarial Attack with Fusion of Polarization for Unmanned Scenes

With the rise of artificial intelligence, the emergence of unmanned vehicles can alleviate traffic congestion and reduce the risk of traffic accidents, in which image recognition has become one of the key technologies. Yet, with the advent of the concept of adversarial examples, many works have proved that the existence of adversarial examples has huge hidden danger in the field of scene recognition. Currently, in unmanned scene recognition, polarization images are widely used because they can robustly describe important physical properties of the object. However, most of the researches on adversarial examples are based on RGB images, and few people studied polarization-based imaging. Therefore, this paper proposes an adversarial attack with fusion of polarization for unmanned scenes. Theoretically, we analyze that polarization images have better effects on adversarial example attacks than RGB images. Experimentally, we evaluate the performance of the proposed model by generating adversarial examples attack scene recognition classification model. The experiment results show that compared with RGB images, polarization images are less vulnerable to attack and have better effects on robustness, which can improve the security of unmanned scenes. And it can reduce the successful attack rate of adversarial examples by up to 9.4%.