Enhancement Robustness of Breast Cancer Models Against Adversarial Attacks

Abstract

 Adversarial attacks pos a critical threat to the reliability of deep learning models, especially in medical imaging, where small pixel-level perturbations can make severe diagnostic misclassifications. This research discusses the vulnerability of a breast cancer histopathology classification model to adversarial attacks and proposes a robust defense framework. Using the BreaKHis 400X dataset, a DenseNet121-based transfer learning model achieved 87.16% accuracy on clean data. But under a Fast Gradient Sign Method (FGSM) attack with ε = 0.05, accuracy decreased to 19.45%, with an attack success rate of 80.55%. To handle these issues, a highly accurate defense model was developed, integrating adversarial training with a denoising module and adversarial detection. The proposed model significantly improved robustness that raising sensitivity to 96.75% and specificity to 90.91% on clean data, while dramatically reducing false negatives. The results show that combining adversarial training with targeted preprocessing can effectively enhance model resilience by offering a practical pathway toward more secure and reliable deep learning systems in clinical settings.