Scholarlink

In recent years, machine learning has gained applications in various fields, but it has its own limitations as well. These limitations can be in the form of the computational power of the machine in use, vulnerabilities against attacks, etc. With the onset of adversarial attacks, the deployment of machine learning/deep learning models has started facing a serious threat of becoming unreliable. Deep learning models have found their way into healthcare in the fields of prognosis to diagnosis, but because of carefully crafted images to cheat these systems, there is a lot of concern about the practical implementation of these models due to the very sensitive nature of the field of medical science. In this paper, we study the consequence of an adversarial attack, the Fast Gradient Sign Method (FGSM), on the Covid-19 X-Ray image dataset, and its effect on the classification accuracy. Different Convolutional Neural Networks (CNNs), including Simple CNN, InceptionV3, InceptionResNetV2, and MobileNet, have been used to determine the performance degradation that is caused by the FGSM. The accuracies drop from a range of 86–92% to 33–39% which is a very significant amount and discourages the use of machine learning without the deployment of some defense mechanisms, like adding adversarial examples during the training of models, which would make the whole system robust or help mitigate these attacks.

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

Assessing the Impact of FGSM Adversarial Attack on Convolutional Neural Networks in X-ray Images

  • 章节

  • 发表日期：

    2026年

查看全文
  • Harshita,
  • Kuldeep Kumar,
  • Sushil Kumar

摘要

In recent years, machine learning has gained applications in various fields, but it has its own limitations as well. These limitations can be in the form of the computational power of the machine in use, vulnerabilities against attacks, etc. With the onset of adversarial attacks, the deployment of machine learning/deep learning models has started facing a serious threat of becoming unreliable. Deep learning models have found their way into healthcare in the fields of prognosis to diagnosis, but because of carefully crafted images to cheat these systems, there is a lot of concern about the practical implementation of these models due to the very sensitive nature of the field of medical science. In this paper, we study the consequence of an adversarial attack, the Fast Gradient Sign Method (FGSM), on the Covid-19 X-Ray image dataset, and its effect on the classification accuracy. Different Convolutional Neural Networks (CNNs), including Simple CNN, InceptionV3, InceptionResNetV2, and MobileNet, have been used to determine the performance degradation that is caused by the FGSM. The accuracies drop from a range of 86–92% to 33–39% which is a very significant amount and discourages the use of machine learning without the deployment of some defense mechanisms, like adding adversarial examples during the training of models, which would make the whole system robust or help mitigate these attacks.