Robust Deep Learning Model for Adversarial Defense in Medical Image Diagnosis

The rapid advancement and widespread deployment of Deep Learning (DL) algorithms across diverse domains have transformed all facets of daily life. In the healthcare sector, advanced learning techniques have demonstrated extraordinary potential, proving effective in various applications, including radiology, pathology, dermatology, and ophthalmology. These innovative learning models have been successfully deployed for medical image analysis; however, the inherent risk of adversarial attacks remains a critical threat. These attacks, designed to introduce subtle yet purposeful modifications to input data, pose a serious challenge to the stability and reliability of such frameworks, particularly in sensitive fields like healthcare. Such attacks can compromise a model’s decision-making, potentially leading to misdiagnoses, delayed treatments, unnecessary procedures, and increased risks to patient safety. This research aims to develop an advanced learning framework for both single-label and multi-label classification tasks using open-access datasets in radiology and pathology. To enhance the model’s resilience against adversarial attacks, we incorporated adversarial training during its development phase. We then employed a pre-trained CheXNet model and evaluated its susceptibility to three targeted adversarial techniques: the perturbation-based method, projected gradient optimization, and the sign gradient-based method. The results indicate that it significantly improves the model’s ability to withstand adversarial attacks and maintains strong performance even under adversarial conditions. This study highlights the importance of developing robust deep learning models for medical image analysis that can withstand such threats. The defensive strategies we have implemented demonstrate significant potential in enhancing the security and robustness of deep learning models in medical applications.