Adversarial Medical Image with Hierarchical Feature Hiding

Area of Science:

• Medical Imaging
• Deep Learning Security
• Computer Vision

Background:

• Deep learning models for medical images are susceptible to adversarial examples (AEs), creating security risks in clinical settings.
• Conventional adversarial attacks (e.g., PGD) create detectable outliers in the feature space, allowing for effective reactive defenses.

Purpose of the Study:

• To investigate the characteristics of conventional medical AEs and reassess the reliability of existing reactive defenses.
• To develop a novel method for generating more stealthy adversarial examples in medical imaging.

Main Methods:

• Theoretical proof showing conventional attacks optimize features in a fixed direction, creating outlier representations.
• A stress test comparing the vulnerability of medical images to natural images.
• Proposal of Hierarchical Feature Constraint (HFC), an add-on to white-box attacks to embed AEs within the target feature distribution.

Main Results:

• Medical images exhibit a unique vulnerability exploitable for hiding AEs.
• The proposed HFC method successfully hides adversarial features, bypassing state-of-the-art medical AE detectors.
• HFC outperforms competing adaptive attacks in evading detection across diverse medical datasets (2D/3D, multiple modalities).

Conclusions:

• Conventional reactive defenses against medical AEs are insufficient due to the distinct characteristics of these attacks.
• The HFC method demonstrates a significant advancement in generating stealthy adversarial examples for medical imaging.
• Findings underscore the need for developing more robust and adaptive defense strategies against sophisticated adversarial attacks in healthcare AI.