RPG-Diff: Precise Adversarial Defense Based on Regional Positioning Guidance

Deep learning models have achieved unprecedented success in various visual tasks. However, they are highly susceptible to adversarial perturbations that are imperceptible to humans. The presence of this risk significantly hinders the deployment of deep learning models in real-world scenarios, especially in safety-critical contexts. Input reconstruction has emerged as an effective defense mechanism against adversarial examples, capable of countering both known and unknown adversarial attacks. Nonetheless, methods based on input reconstruction often suffer from slow training speeds and unsatisfactory reconstruction quality, limiting their practical applicability.To address these issues, we propose a region localization-guided input reconstruction defense method, named RPG-Diff. This approach utilizes image segmentation algorithms to locate key regions, which then guide the reconstruction process of the diffusion model. By leveraging information from these key regions, the diffusion model can more effectively eliminate adversarial perturbations while maintaining consistency between the reconstructed output and the original input. This significantly reduces the impact of adversarial perturbations on model accuracy. Experimental results demonstrate that, compared to existing methods, our approach exhibits superior defensive performance against various adversarial attacks and surpasses other preprocessing-based defense methods in terms of robust accuracy and average generation time.