A Survey of Defense Methods Against Adversarial Examples

With the development of artificial intelligence technology, to our surprise, deep neural networks have made a series of breakthroughs and achieved optimal results in lots of machine learning tasks, while the image field is the most prominent. Although deep learning technology has brought great convenience to our lives, the security problems of artificial intelligence have gradually become prominent. But deep neural networks are very fragile, and usually a small perturbation can have a serious impact on the neural network, especially adversarial examples. As a small imperceptible perturbation, it can cause malicious effects on the results of image classification. What can we do to protect clean examples from perturbations has become the most important and practical research direction in the current image field. A large number of scholars have begun to invest in adversarial examples research, especially in defense, the achievements are more abundant, and related research results have also shown a blowout state. Based on the research status of recent years, this paper first explains the concept of adversarial examples, organizes the current attack methods, and then summarizes different defense methods according to the criteria of detection defense and robust defense. Improvements are discussed to help further development of adversarial defense of images.