Curvature-Regularized Robust Medical Image Classification
Novel deep learning framework combining CURE loss, depthwise separable & dilated convolutions, and attention mechanisms — achieving only 5.6% accuracy drop under adversarial attacks.
Problem
Medical image classification models are vulnerable to adversarial perturbations, raising safety concerns in clinical deployment. Standard training fails to produce robust models under input noise.
Action
Developed MedMamba framework applying curvature regularization (CURE) to control Hessian eigenvalues for enhanced robustness. Designed a dual-branch SS-Convd-SSM module combining dilated and depthwise separable convolutions with attention for precise feature extraction. Validated robustness through FGSM and PGD adversarial attacks.
Result
Under adversarial attack, the model showed only a 5.6% drop in average accuracy and 4.1% drop in AUC — demonstrating strong robustness compared to baseline methods. Published at ICCGIV 2024.
Learnings
Gained expertise in adversarial robustness, curvature regularization theory, and designing neural architectures that balance efficiency with resilience to input perturbations.