ULAN: A Universal Local Adversarial Network for SAR Target Recognition Based on Layer-Wise Relevance Propagation

Recent studies have proven that synthetic aperture radar (SAR) automatic target recognition (ATR) models based on deep neural networks (DNN) are vulnerable to adversarial examples. However, existing attacks easily fail in the case where perturbations cannot be fully fed victim models. We call this situation perturbation offset. Moreover, since background clutter takes up most of area SAR images and has low relevance results, fooling with global is quite inefficient. This paper proposes a semi-white-box attack network called Universal Local Adversarial Network (ULAN) generate universal (UAP) for regions images. In proposed method, we calculate model’s attention heatmaps through layer-wise propagation (LRP), which used locate high results. particular, utilize generator U-Net learn mapping from noise UAPs craft examples by adding generated local regions. Experiments indicate method effectively prevents offset achieves comparable performance conventional perturbing only quarter or less image areas.