RDTD: Adversarial Input Detector Using Reverse Diffusion Transformation

Deep neural networks (DNNs) have achieved performance beyond humans in visual recognition tasks. However, DNNs are susceptible to adversarial examples spoofing, where an attacker adds malicious perturbations to an image causing the neural network to output incorrect results, and this attack often has a high degree of similarity to the original image at the visual level. This paper proposes a reverse diffusion transform detector(RDTD) to defend against such malicious input. We have conducted extensive experiments on MNIST, CIFAR10, and ImageNet datasets using different models, and the results show that our proposed detection framework is effective in detecting adversarial attacks due to the existing defenses. and also shows that our detection framework has better adaptability. Previous defense frameworks do not consider the effectiveness shown in the face of novel digital attacks proposed in recent years. We evaluate the detection effectiveness of the proposed framework in producing adversarial samples from a data distribution perspective, showing that our approach still has high performance in the face of advanced attacks, and demonstrating that our proposed framework can still provide effective defense for deep neural networks against novel attacks.