The rapid growth of machine learning (ML) has produced remarkable achievements across varied research areas. However, this progress has brought forth a novel challenge known as adversarial attacks. An ML approach called Adversarial Machine Learning (AML) is used to trick models using hostile input, frequently leading to system failures. This technique may be used for different data and model sets as it can be utilized to alter input data, interpret data incorrectly, and cause the system to inaccurately identify data when the application is being implemented. The integrity and performance of ML systems can be seriously jeopardized by adversarial attacks, which include targeted, non-targeted, white-box, black-box, and gray-box attacks, etc. While certain defenses, such as adversarial training, robust optimization frameworks, and adversarial prediction problems, attempt to improve model robustness, real robustness is still difficult to achieve because of large computing costs and possible accuracy trade-offs. Although machine learning’s explosive growth has produced many technological advances, it also raises questions about how susceptible it is to hostile attacks. In addition to describing AML concepts, current research, attack methods, and practical applications, this article addresses the security vulnerabilities associated with malevolent attackers influencing and eluding ML systems.