Adversarial robustness of dental panoramic radiograph classifiers against gradient-based perturbations

Introduction: Deep learning classifiers for dental radiographs are increasingly deployed, but their robustness to adversarial perturbations is unmeasured. Methods: A frozen ResNet-18 backbone and a linear head were trained on 182 annotated teeth from the DENTEX dataset and attacked with the fast gradient sign method and projected gradient descent at budgets from 1/255 to 16/255, comparing a baseline model, an adversarially trained model and a median-filter defense. Results: The baseline classified 63.6% of held-out teeth correctly. Both attacks flipped every correctly classified tooth at a budget of 1/255. The fast gradient sign method fell to 77.1% attack success at 16/255, whereas projected gradient descent remained at 100%. Neither adversarial training nor the median filter reduced attack success, and adversarial training lowered clean accuracy to 54.5%. Conclusions: A transfer-learning dental radiograph classifier with a frozen backbone and linear head is catastrophically vulnerable to adversarial perturbations, and the simple defenses tested do not restore robustness.