Enhancing resilience against adversarial attacks in medical imaging using advanced feature transformation training

Abstract

This study presents a machine learning-driven defense mechanism against adversarial attacks, specifically tailored for medical imaging applications. This mechanism utilizes feature transformation through transfer learning, leveraging a fine-tuned ResNet152V2 network trained on original medical images. To enhance the model's robustness, we apply efficient adversarial training on transformed features extracted from both original and adversarial images. Additionally, we integrate Principal Component Analysis (PCA) to reduce feature dimensionality, optimizing the adversarial training process. When evaluated on Chest X-ray datasets, focusing on pneumonia and normal cases, the proposed mechanism demonstrated strong resilience against imperceptible attacks while maintaining a performance retention rate above 90 %. These results show the potential of the proposed mechanism to enhance the reliability and security of CNN-based medical imaging systems in practical, real-world settings.