The paper introduces Defense-GAN, a novel framework that leverages generative models to protect deep neural networks from adversarial attacks. Defense-GAN is trained to model the distribution of unperturbed images and, at inference time, projects adversarial inputs onto the range of the generator's output to reduce the adversarial perturbations. This projected output is then fed to the classifier. The method is effective against both white-box and black-box attacks and can be used with any classification model without modifying the classifier structure or training procedure. Empirical results on benchmark datasets show that Defense-GAN consistently outperforms existing defense strategies, demonstrating its robustness and effectiveness against various attack methods. The code for Defense-GAN is publicly available at https://github.com/kabkabm/defensegan.The paper introduces Defense-GAN, a novel framework that leverages generative models to protect deep neural networks from adversarial attacks. Defense-GAN is trained to model the distribution of unperturbed images and, at inference time, projects adversarial inputs onto the range of the generator's output to reduce the adversarial perturbations. This projected output is then fed to the classifier. The method is effective against both white-box and black-box attacks and can be used with any classification model without modifying the classifier structure or training procedure. Empirical results on benchmark datasets show that Defense-GAN consistently outperforms existing defense strategies, demonstrating its robustness and effectiveness against various attack methods. The code for Defense-GAN is publicly available at https://github.com/kabkabm/defensegan.