Scholarlink

<p>Deep and machine learning models have become pivotal in medical image analysis, especially for diagnosing COVID-19 using X-rays and CT scans. While these models, including transfer learning-based approaches, have achieved high accuracy, they remain highly vulnerable to adversarial attacks, which can manipulate input data and cause misclassification, posing critical risks in clinical applications. This study introduces a novel approach to addressing this issue by systematically evaluating the impact of adversarial attacks on COVID-19 diagnosis models built with two leading architectures, VGG-16 and DenseNet-121, using the Fast Gradient Sign Method (FGSM). The FGSM attack causes a dramatic drop in accuracy, reducing VGG-16’s accuracy from 95.12 to 9.97% and DenseNet-121’s from 96.51 to 10.13%. To counter these vulnerabilities, we propose a novel defense mechanism that combines adversarial training with Gaussian noise data augmentation, a dynamic approach that generates perturbations across various epsilon values during the training phase. This innovative method significantly enhances model robustness, restoring accuracy to over 92% on adversarial examples. These findings emphasize the need for strong defense mechanisms in deep learning models for COVID-19 diagnosis, ensuring reliability and security against adversarial threats in clinical environments.</p>

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

Threats to medical diagnosis systems: analyzing targeted adversarial attacks in deep learning-based COVID-19 diagnosis

  • 期刊论文

  • 发表日期：

    2025年2月17日

查看全文
  • Sheikh Burhan Ul Haque,
  • Aasim Zafar,
  • Sheikh Riyaz Ul Haq,
  • Sheikh Moeen Ul Haque,
  • Mohassin Ahmad,
  • Khushnaseeb Roshan

摘要

Deep and machine learning models have become pivotal in medical image analysis, especially for diagnosing COVID-19 using X-rays and CT scans. While these models, including transfer learning-based approaches, have achieved high accuracy, they remain highly vulnerable to adversarial attacks, which can manipulate input data and cause misclassification, posing critical risks in clinical applications. This study introduces a novel approach to addressing this issue by systematically evaluating the impact of adversarial attacks on COVID-19 diagnosis models built with two leading architectures, VGG-16 and DenseNet-121, using the Fast Gradient Sign Method (FGSM). The FGSM attack causes a dramatic drop in accuracy, reducing VGG-16’s accuracy from 95.12 to 9.97% and DenseNet-121’s from 96.51 to 10.13%. To counter these vulnerabilities, we propose a novel defense mechanism that combines adversarial training with Gaussian noise data augmentation, a dynamic approach that generates perturbations across various epsilon values during the training phase. This innovative method significantly enhances model robustness, restoring accuracy to over 92% on adversarial examples. These findings emphasize the need for strong defense mechanisms in deep learning models for COVID-19 diagnosis, ensuring reliability and security against adversarial threats in clinical environments.