Universal adversarial defense in remote sensing based on pre-trained denoising diffusion models

Weikang Yu, Yonghao Xu, Pedram Ghamisi

Abstract: Highlights•Introduces UAD-RS for universal adversarial defense on RS adversarial examples.•Mitigates adversarial perturbations with unified pre-trained diffusion models.•Implements an ANLS mechanism optimizing noise levels with FID ranking.•Demonstrates effectiveness for scene classification and semantic segmentation.•Outperforms SOTA approaches defending against seven adversarial perturbations.