The Evolution of Adversarial Attacks in Deep Learning: A Holistic Survey

DNNs are vulnerable to subtle but malicious input perturbations, which compromising their predictions and posing serious security issues, which called adversarial attacks This article demonstrates the effectiveness of this kind of action on DNNs, shedding light on dangers related to extract data that could be unauthorized in sensitive domains like audio and text. The adversarial techniques discussed include advanced methods like Project Gradient Descent (PGD) and Basic Iterative Method (BIM), which enhance FGSM by iteratively applying perturbations. Graph Neural Networks (GNNs) are vulnerable too, where attackers transform graph structures and node features that cause the model to degrade accuracy. It reviews and evaluates defense strategies such as adversarial training and partitioned batch normalization, while they contend with a significant attack problem with the gradient based adversaries. The findings highlight the need for regulatory frameworks and protections around data integrity, noting that given the heavy reliance on data for systems of increasingly critical applications, including self-driving vehicles and medical imaging systems, robust defenses must be developed.