Although deep learning has developed rapidly in recent years, the existence of adversarial examples casts a shadow over its future. Deep neural networks (DNNs) can be unreliable when facing these carefully crafted adversarial examples, limiting their real-world applications. While adversarial examples usually lack robustness and tend to lose effect when the viewing angle is changed. Moreover, it is impossible to control the angle at which adversarial examples are seen by the airborne detector. Therefore, we propose a multi-view robust adversarial attack by introducing perspective transformation and feature similarity loss to mitigate the negative effect that viewing angle changes bring to the features. We evaluated our adversarial patch on our multi-view test dataset and the results show that our proposed method reduced the mAP by 73.4%. Furthermore, compared to other methods, our approach performs best in the data with low depression angles, demonstrating its excellent multi-view robustness.
|