Scholarlink

Adversarial attacks pose a real challenge to artificial intelligence in medical imaging, as the very small changes caused by these attacks, which are indistinguishable to the naked eye, can lead to misclassification of these images, resulting in incorrect diagnoses and putting patient safety at risk. In this study, we examine how three different AI architectures—DenseNet-121 (CNN), Vision Transformer (ViT), and hybrid ConViT—respond to common Adversarial attacks: Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), and Carlini–Wagner (CW) attacks. While all models perform well on clean medical images, adversarial perturbations reveal important differences: CNNs show moderate resilience to simple attacks, while ViTs are highly sensitive to stronger repeated attacks, and ConViT hybrid models achieve a balance by effectively combining local and global features. From a clinical perspective, these adversarial effects can increase the likelihood of misdiagnosis, highlighting the urgent need for robust AI designs. Our results suggest that hybrid architectures, which combine transfer and transformation features, may provide a safer and more reliable approach to deploying AI in healthcare.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Adversarial Robustness of CNN, Transformer, and Hybrid Architectures in Medical Imaging AI

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Enshirah Altarawneh,
  • Esraa Alshdaifat

摘要

Adversarial attacks pose a real challenge to artificial intelligence in medical imaging, as the very small changes caused by these attacks, which are indistinguishable to the naked eye, can lead to misclassification of these images, resulting in incorrect diagnoses and putting patient safety at risk. In this study, we examine how three different AI architectures—DenseNet-121 (CNN), Vision Transformer (ViT), and hybrid ConViT—respond to common Adversarial attacks: Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), and Carlini–Wagner (CW) attacks. While all models perform well on clean medical images, adversarial perturbations reveal important differences: CNNs show moderate resilience to simple attacks, while ViTs are highly sensitive to stronger repeated attacks, and ConViT hybrid models achieve a balance by effectively combining local and global features. From a clinical perspective, these adversarial effects can increase the likelihood of misdiagnosis, highlighting the urgent need for robust AI designs. Our results suggest that hybrid architectures, which combine transfer and transformation features, may provide a safer and more reliable approach to deploying AI in healthcare.