Siamese Denoising Autoencoders for Enhancing Adversarial Robustness in Medical Image Analysis

Deep learning models have achieved groundbreaking results in computer vision; however, their vulnerability to adversarial examples persists. Adversarial examples, generated by adding minute perturbations to images, lead to misclassification and pose serious threats to real-world applications of deep learning models. This paper proposes a simple, powerful, and efficient adversarial defense method: a Siamese network-based Denoising Autoencoder (Siamese-DAE). This method addresses the classification accuracy reduction caused by the denoising process. Experiments on Chest X-ray, Brain MRI, Retina, and Skin images, using FGSM, PGD, DeepFool, CW, SPSA, and AutoAttack adversarial algorithms, demonstrate that the Siamese-DAE, trained to remove noise, effectively eliminates perturbations, leading to improved classification accuracy compared not only to the standard classification model but also to relevant denoising defense models.