AI Driven Medical Diagnostics with Improved Robustness to Adversarial Perturbations

This research introduces a robust AI framework for automated lung cancer classification using CT images, addressing the vulnerability of deep convolutional neural network (CNN) models to adversarial attacks. A baseline CNN trained on SMOTE-balanced clean data achieved a high validation accuracy of 98.91% under clean conditions. However, when evaluated under Projected Gradient Descent (PGD) attack, the model’s accuracy dropped drastically to 4.00%, with a complete misclassification of normal cases. To mitigate this, we implemented a defensive distillation pipeline, where a high-temperature teacher model generated soft labels for training a student model. The student was trained on a hybrid dataset composed of 40% clean samples (with soft labels) and 60% adversarial examples (PGD and FGSM), enhanced through label smoothing and symmetric cross-entropy loss. The robust student model preserved high clean-data performance at 97.82% accuracy and significantly improved robustness under PGD attack, reaching 96.36% accuracy, a 24-fold improvement over the baseline under attack. These results validate the effectiveness of adversarial training and distillation in developing resilient diagnostic AI systems.