Adversarial Attacks and Defenses in Machine Learning Models

Adversarial attacks carry a major threat to the reliability and safety of machine learning (ML) models. These types of attacks provide inputs that exploit the shortcomings of machine learning systems, thereby producing erroneous predictions by means of such algorithms. This paper offers a thorough overview of hostile attacks along with their consequences in the real world as well as the developed defensive strategies meant to either lower or stop them. Using case studies and empirical data, we look at a broad spectrum of adversarial assaults including physical attacks, black-box attacks, and white-box attacks. We also discuss the benefits and drawbacks of each of the most sophisticated defensive strategies already in use— adversarial training, input preprocessing, and randomised smoothing. At last, we have found the open challenges in the field of adversarial machine learning as well as future directions of study possibilities. The intended audience for this essay, which aims to be a useful tool for researchers and practitioners interested in knowing and handling the problems adversarial attacks bring about, is themselves.