Systemic Adversarial Vulnerability in Deep Learning Classifiers: A Cross-Model Experimental Analysis

Adversarial attacks demonstrate the existence of severe weaknesses in deep learning image classifiers due to misclassification, caused by a small, precisely-designed input per- turbation. This work aims to experimentally test the adversarial robustness and the extent to which adversarial vulnerability can be model-dependent or systemic. There are two controlled experiments. First, on a pretrained ResNet-18 model, Fast Gra- dient Sign Method (FGSM), Projected Gradient Descent (PGD) and Carlini-Wagner (CW) attacks are applied to discuss the impact of attack sophistication. Second, cross-model robustness is compared by means of eight pretrained image classification models to evaluate FGSM for cross-model robustness. The findings indicate that optimization-based attacks are much better than single-step attacks and more complex architectures are more robust and lightweight models are more sensitive. These results suggest that adversarial vulnerability is architecture-dependent and exists irrespective of a model. The paper also discusses defense and recovery processes and points out how the principles of Responsible AI can be used to construct high-quality and trustworthy adversarial defenses.