<![CDATA[Deep learning, especially Convolutional Neural Network (CNN) architectures, has significantly improved medical image analysis for predicting lung diseases through chest X-ray (CXR) images, including pneumonia and COVID-19. However, despite achieving high diagnostic precision, CNN models remain highly susceptible to adversarial attacks, defined as small, visually imperceptible alterations optimized to exploit non-linear decision boundaries that cause high-confidence mispredictions. This vulnerability presents a critical concern in clinical settings, where deterministic diagnostic errors directly compromise patient safety. This paper systematically implements white-box adversarial attacks to quantify the resilience of CNN models in multi-class CXR image classification. This paper utilizes the COVID-19 Radiography Dataset, comprising four diagnostic categories: COVID-19, Lung Opacity, Normal, and Viral Pneumonia. A DenseNet-121 architecture was employed for feature extraction, and the trained model was subsequently subjected to Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) attacks under varying Lāā-bounded epsilon settings. The empirical experiments reveal three critical findings: 1) The implementation of sub-pixel adversarial attacks causes severe performance degradation, where the PGD attack constrained at an epsilon of 0.1/255 reduced the global model accuracy from a baseline of 95.42% to 25.32%; 2) Iterative attacks (PGD) represent the absolute worst-case scenario for model reliability by efficiently discovering high-dimensional manifold gaps, whereas the model demonstrates relative resilience to linear, single-step FGSM perturbations; and 3) Gradient-weighted Class Activation Mapping (Grad-CAM) analysis verifies that this performance collapse is associated with a deterministic semantic shift, displacing the model's spatial attention from clinically relevant pulmonary regions toward spurious background noise. In conclusion, this paper empirically proves that despite exhibiting high accuracy on clean data, unprotected CNNs remain fundamentally unsafe for autonomous clinical deployment due to their acute vulnerability to gradient-based perturbations, necessitating the future integration of robust adversarial training frameworks]]>
