Adversarial Defense in Aerial Detection

Yuwei Chen, Shiyong Chu; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2023, pp. 2306-2313

The excellent performance of artificial intelligence algorithms in target detection greatly improves the efficiency of detection. However, this alternative to human processing of image information faces many challenges, one of which is adversarial examples (AE). For aerial detection, it is a function widely used in many fields to obtain detection pictures of optical, infrared, and synthetic aperture radars (SAR) from high altitudes to identify ground targets. But in the current research results, optical sensors, infrared sensors, and SAR will be attacked by adversarial patches and perturbation. When these attacks exist, it is risky to let intelligent algorithms perform aerial detection. This paper will focus on the characteristics of each detection mode and propose Adaptive Defense Pipeline (ADP) in addition to improving algorithm robustness through training. According to different weather conditions, the ADP sets the weight coefficients of the detection results of multiple sensors to synthesize the detection results, and on this basis, the second confirmation is added. At the same time, we compare the traditional aerial detection results of a single sensor with the weighted results using ADP and verify that the proposed method could indeed improve the efficiency of aerial detection using artificial intelligence algorithms in an adversarial environment.