Abstract
Deep learning models have achieved remarkable success in medical image classification; however, they remain highly susceptible to adversarial perturbations that can drastically alter diagnostic predictions. This study implements and evaluates a curriculum-based adversarial training strategy aimed at improving model robustness and reliability in clinical image analysis. A DenseNet-121 architecture was trained on clean medical datasets to establish baseline diagnostic performance. Subsequently, adversarial examples generated using five gradient-based attacks, namely FGSM, BIM, PGD, MIFGSM, and APGD, were progressively incorporated through a curriculum schedule that gradually increased the ratio of perturbed samples. This incremental exposure enabled the model to adapt to adversarial noise while maintaining stable recognition of disease-relevant features. Experiments were conducted on two imaging domains: chest X-rays for pneumonia detection and multi-class kidney CT image classification. The clean baseline models exhibited severe performance degradation when tested on mixed datasets containing both clean and adversarial images. In contrast, the adversarially trained models demonstrated substantially higher resilience, achieving approximately 10–12 percentage points greater accuracy on mixed data compared to the clean models, while retaining comparable performance on clean datasets. These results confirm that curriculum-based adversarial training enhances robustness without compromising diagnostic fidelity, offering a reproducible pathway toward trustworthy and deployment-ready AI systems for medical imaging applications..