Evaluation and Analysis of Robustness of Adversarial Examples Attacks in Deep Neural Networks

摘要

Neural networks have revolutionized the field of artificial intelligence. They have given rise to several applications in various areas. However, it has been shown that they have flaws that could be leveraged by an attacker to perform an adversarial examples attack. Several studies have attempted to design defensive mechanisms that would ensure the security of neural networks. Nevertheless, these mitigation techniques remain insufficient to entirely address all the vulnerabilities that may reside in these architectures. In this article, we will extensively address the security of neural networks. We will begin with a study of the different approaches towards the security of neural networks. Then, we will detail significant sources of vulnerabilities in neural networks. Afterward, we will examine the theory of adversarial examples attack as well as the optimization problem related to them. Thereafter, we will conduct a comparative study of the most common mitigation techniques in the literature. Finally, we will propose a framework devoted to assessing adversarial examples robustness.