Robust Defense Against Adversarial Attacks with Defensive Preprocessing and Adversarial Training

Adversarial attacks, achieved through carefully crafted perturbations, exploit vulnerabilities in deep learning models, posing serious risks. In this study, we present an integrated defensive preprocessing and adversarial training pipeline as a unified and effective defense against multiple attacks targeting image classification models. Defensive preprocessing introduces random noise to mitigate harms from attacks, employing average pooling and denoising to effectively reduce their impact. Super-resolution is used to recover lost details post-pooling. Adversarial training generates additional samples under various attacks, enabling the model to learn and recognize attacked images. Our findings demonstrate that the combination of defensive preprocessing and adversarial training effectively defends against multiple unknown attack types, requiring fewer training types and samples. The proposed defensive method is versatile, with minimal computational demands and without the need to retrain the underlying model. Furthermore, it can be extended to handle new, unseen attack types with training on just a single attack type.