Universal Adversarial Examples in Remote Sensing: Methodology and Benchmark

Deep neural networks have achieved great success in many important remote sensing tasks. Nevertheless, their vulnerability to adversarial examples should not be neglected. In this study, we systematically analyze the Universal Adversarial Examples Remote Sensing (UAE-RS) data for first time, without any knowledge from victim model. Specifically, propose a novel black-box attack method, namely, Mixup-Attack, and its simple variant Mixcut-Attack, data. The key idea of proposed methods is find common vulnerabilities among different by attacking features shallow layer given surrogate Despite simplicity, can generate transferable that deceive most state-of-the-art deep both scene classification semantic segmentation tasks with high rates. We further provide generated universal dataset named UAE-RS, which provides samples field. hope UAE-RS may serve as benchmark helps researchers design strong resistance toward attacks Codes are available online (https://github.com/YonghaoXu/UAE-RS).