Reducing Perturbation of Adversarial Examples via Projected Optimization Method

Deep neural networks are vulnerable to the adversarial example. So far, the primary way of generating the adversarial example is comparing the success rates of the adversarial attack. However, the distance between the made example and the original example is also an essential indicator. In this paper, it is demonstrated that the optimization algorithm could reduce the perturbation of adversarial example generated by using the extremum loss function to obtain the perturbation. This paper introduces the OPA optimization algorithm and uses it to find the best advantage on the model decision boundary as the adversarial example. This paper tests four attack methods FGSM, BIM, MI-FGSM and CW, and measures the perturbation between the original sample and the generats sample by the Euclidean distance. And it is found that the noise of the sample image is significantly reduced by OPA optimization. It should be set in 12-point font size.