Protecting Sar Imagery From Adversarial Attack Using A Diffusion Model

Efficient wide-area surveillance of the maritime environment requires automated detection and classification techniques. Deep learning can be exploited for this purpose, but is susceptible to adversarial attacks that can cause misclassifications. Diffusion models have previously been shown to improve robustness to adversarial perturbation in electro-optic imagery. The aim of this work is to investigate how this adversarial defence technique protects against white-box attacks on a high-resolution Synthetic Aperture Radar (SAR) dataset. The benefits of the diffusion model are demonstrated using a maritime SAR dataset, and include an improvement in the classification performance and a reduction of the sea clutter.