Adversarially Enhanced Learning (AEL): Robust lightweight deep learning approach for radiology image classification against adversarial attacks

Abstract

Deep learning models perform well in medical image classification, particularly in radiology. However, their vulnerability to adversarial attacks raises concerns about robustness and reliability in clinical applications. To address these concerns, a novel approach for radiology image classification, referred to as Adversarially Enhanced Learning (AEL) has been proposed. This approach introduces a novel deep learning model ConvDepth-InceptNet designed to enhance the robustness and accuracy in radiology image classification through three key phases. In Phase 1, adversarial images are generated to deceive the classifier using the proposed model initially trained for classification. Phase 2 entails re-training the model with a mix of clean and adversarial images, improving its robustness by functioning as a discriminator for both types of images. Phase 3 refines adversarial images with Total Variation Minimization (TVM) denoising before classification by re-trained model. Pre-attack analysis with VGG16, ResNet-50, and XceptionNet achieved 98% accuracy with just 10,946 parameters. Post-attack analysis subjected to attacks such as Fast Gradient Sign Method, Basic Iterative Method, and Projected Gradient Descent, yields an average adversarial accuracy of 94.8%, with standard deviation of 1.6%, and an attack success rate of 3.3%. Comparative analysis with ResNet50, VGG16, and InceptionV3 indicates minimal performance drops. Furthermore, post-defense analysis shows that the adversarial images refined with TVM denoising are evaluated with re-trained model, achieving an outstanding ac- curacy of 98.83%. The combination of denoising techniques (Phase 3) and robust re-training (Phase 2) enhances robustness by providing a layered defense mechanism. The analysis validates the robustness of this approach.