Year: 2026
Type: dissertation
Abstract: Advances in deep learning have revolutionized lung disease diagnosis through chest X-ray (CXR) images, particularly since the COVID-19 pandemic. Convolutional Neural Network (CNN) architectures such as DenseNet-121 demonstrate high accuracy in distinguishing viral pneumonia, lung opacity, and normal cases. However, recent studies reveal that these models possess fundamental weaknesses against iteratively designed adversarial attacks. This study aims to analyze the robustness of the DenseNet-121 model against the Projected Gradient Descent (PGD) attack in a four-class classification task on COVID-19 chest X-ray images. The methodology involves training the model using transfer learning on the COVID-19 Radiography Dataset (21,165 images), followed by testing with PGD attacks at various epsilon levels (0.1/255 to 0.3/255). The results show that at epsilon 0.1/255, global model accuracy drops drastically from 95.42% to 25.32%, with an Attack Success Rate (ASR) reaching 74.68%. The most severe degradation occurs in the Lung Opacity class (F1-score drop of 88.90%) and the COVID-19 class (F1-score drop of 82.99%). Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations reveal the phenomenon of semantic decoupling, where the model's attention shifts from diseased lung regions toward background artifacts. In conclusion, the iterative nature of PGD attacks effectively exploits gaps in high-dimensional decision manifolds, rendering unprotected CNN models unsuitable for direct deployment in attack-prone clinical environments.
Author: Farma, Elga Putri Tri
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7219288436
