A Feature Space-Restricted Attention Attack on Medical Deep Learning Systems.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Healthcare
• Cybersecurity in Medicine

Background:

• Deep neural networks (DNNs) excel in medical image analysis but are vulnerable to adversarial attacks.
• These attacks, minor input disruptions, pose significant security risks in the health information economy.
• Understanding these vulnerabilities is crucial for deploying AI in clinical settings.

Purpose of the Study:

• To analyze adversarial attacks on medical deep learning systems from white-box and black-box perspectives.
• To propose a novel, fast adversarial sample generation method for more effective and stealthy attacks.
• To enhance the security and generalization of deep learning models in medical applications.

Main Methods:

• Analysis of adversarial attacks using white-box and black-box methodologies.
• Development of Feature Space-Restricted Attention Attack (FSA^2) using generative adversarial networks (GANs).
• Integration of an attention mechanism to focus perturbations on lesion regions for increased efficiency and invisibility.

Main Results:

• The proposed FSA^2 method generates highly confusing adversarial samples.
• Experiments on three medical image types demonstrate the method's performance and specificity.
• The attention mechanism enhances perturbation efficiency and stealth, making attacks nearly invisible.

Conclusions:

• The study highlights current weaknesses in clinical deep learning system deployment.
• The proposed method offers a way to create more robust and secure medical AI.
• Further research into domain-specific features can improve model generalization and attack resistance.