Natural Images Allow Universal Adversarial Attacks on Medical Image Classification Using Deep Neural Networks with...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Cybersecurity

Background:

• Deep neural networks (DNNs) with transfer learning are used for medical image classification.
• Adversarial vulnerability of DNNs poses risks to clinical diagnosis.
• Medical image datasets are typically unavailable for adversarial attacks due to privacy concerns.

Purpose of the Study:

• To demonstrate adversarial attacks on medical DNNs using natural images.
• To investigate the effectiveness of universal adversarial perturbations (UAPs) generated from natural images.
• To assess the security implications of transfer learning in medical AI.

Main Methods:

• Generating universal adversarial perturbations (UAPs) from natural images.
• Applying UAPs to medical DNN models trained with transfer learning.
• Evaluating UAP performance against non-targeted and targeted attacks.
• Comparing UAP performance with random controls and models trained from random initialization.

Main Results:

• Adversarial attacks are feasible using natural images on medical DNNs with transfer learning.
• UAPs generated from natural images are effective for both non-targeted and targeted attacks.
• The performance of natural image UAPs significantly exceeds random controls.
• Transfer learning introduces a security vulnerability, decreasing diagnostic reliability.

Conclusions:

• Transfer learning in medical DNNs creates a security loophole exploitable by natural image-based adversarial attacks.
• This vulnerability threatens the reliability and safety of computer-aided disease diagnosis.
• While random initialization reduces UAP effectiveness, it does not eliminate the vulnerability, highlighting a significant future security concern.