Energy-Based Adversarial Example Detection for SAR Images

Adversarial examples (AEs) bring increasing concern on the security of deep-learning-based synthetic aperture radar (SAR) target recognition systems. SAR AEs with perturbation constrained to vicinity have been recently in spotlight due physical realization prospects. However, current adversarial detection methods generally suffer severe performance degradation against region-constrained perturbation. To solve this problem, we treated as low-probability samples incompatible clean dataset. With help energy-based models, captured an inherent energy gap between and that is robust changes region. Inspired by discovery, propose detector, which requires no modification a pretrained model. better distinguish AEs, regularization was adopted fine-tune Experiments demonstrated proposed method significantly boosts