Local Pixel Attack Based on Sensitive Pixel Location for Remote Sensing Images

As deep neural networks (DNNs) are widely used in the field of remote sensing image recognition, there is a model security issue that cannot be ignored. DNNs have been shown to vulnerable small perturbations large number studies past, and this risk naturally exists object detection models based on DNNs. The complexity makes it difficult implement adversarial attacks them, resulting current lack systematic research examples recognition. In order better deal with threats recognition may confront provide an effective means for evaluating robustness models, paper takes as goal systematically vanishing against model. To solve problem attack implementation detection, adaptation methods interpolation scaling patch perturbation stacking proposed paper, which realizes classical algorithms. We propose hot restart update strategy joint first second stages two-stage realized through design loss function. For modification cost global pixel being too large, local algorithm sensitive location paper. By searching pixels constructing mask area, good effect achieved. Experimental results show average rate method decreases less than 4% can still maintained above 80%, effectively achieves balance between cost.