A REVIEW OF ADVERSARIAL ATTACK METHODS ON DEEP NEURAL NETWORKS

Because Deep Neural Networks (DNNs) can simulate complicated data representations, they have shown impressive results in a variety of applications, including autonomous systems, natural language processing, and picture classification.  Despite their exceptional performance, DNNs are extremely susceptible to adversarial assaults, in which models provide inaccurate predictions due to carefully planned, undetectable modifications to input data.  Particularly in security-sensitive applications like financial systems, healthcare diagnostics, and autonomous driving, this vulnerability presents serious issues.  This review paper offers a thorough examination of protection mechanisms intended to lessen the threat of various adversarial attack techniques, such as backdoor, data poisoning, black-box, and white-box attacks.  This study highlights important trends, difficulties, and possible avenues for further research in adversarial machine learning by reviewing the body of existing literature.  The study's narrative literature review methodology provides insights into new trends while combining important research findings.

Keywords: Adversarial Attack, Deep Neural Networks, White-box Attack, Black-box Attack, Data Poisoning, Backdoor Attack, Adversarial Defense, Machine Learning Security

Adversarial Attack Deep Neural Networks White-box Attack Black-box Attack Data Poisoning Backdoor Attack Adversarial Defense Machine Learning Security