Robust Adversarial Attack Detection via Generative Adversarial Network With Residual Multi-Layer Aggregation Based Contrastive Loss Function

Abstract

Adversarial attacks in medical imaging refer to subtle modifications to images that mislead diagnostic systems, resulting in inaccurate diagnoses and assessments. These attacks exploit vulnerabilities in image processing, leading to misclassification or altered visual features that often go unnoticed. This raises serious concerns about the security and reliability of medical diagnosis, directly impacting clinical decision-making and patient safety. This research proposes a Generative Adversarial Network with Residual Multi-Layer Aggregation-based Contrastive Loss Function (GRMLA-CLF) to effectively identify adversarial attacks using medical images. In the generator, Residual Multi-Layer Aggregation (RMLA) is incorporated to capture fine-grained information and structural patterns of adversarial attacks, improving the model’s adaptability. The Contrastive Loss Function (CLF) enhances adversarial attack detection by increasing the distance between genuine and adversarial samples, ensuring a clear distinction in latent space, and ensuring distinct representation. This enhances model robustness by reducing sensitivity to small perturbations while preserving significant features necessary for accurate decision-making. The proposed GRMLA-CLF achieves high accuracy rates of 99.81, 99.64, and 98.65% on the ISIC2019, Chest X-ray, and APTOS2019 datasets, respectively, outperforming existing methods like Global Attention Noise (GATN).