Research on Adversarial Sample Defense Method Based on Image Denoising

With the widespread application of deep learning in image processing, the impact of image quality on model performance has become increasingly prominent. This paper proposes an enhanced U-Net network (NS-UNet), which integrates a channel attention mechanism and a blind-spot denoising strategy, to address the challenges posed by both conventional noise and adversarial perturbations. The method introduces an efficient channel attention module (ECA-Net) into the U-Net architecture, enhancing the model's ability to capture key features. Additionally, the blind-spot denoising mechanism allows the model to effectively suppress adversarial perturbations and preserve structural information without relying on reference clean images. Experimental results show that NS-UNet achieves a defense accuracy of 73.5% against PGD attacks on the ImageNet dataset. Furthermore, NS-UNet demonstrates excellent denoising performance on standard datasets such as Kodak, BSD300, and Set14, with an SSIM value of 0.882 on Set14, improving by approximately 3.28% compared to the classic BM3D method. This approach not only enhances image recovery quality in noisy environments but also significantly improves the robustness of deep models, offering an efficient and feasible solution for image denoising and security protection in real-world scenarios.