Universal black-box attacks against a third-party Alzheimer's diagnostic system.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cybersecurity

Background:

• Artificial intelligence (AI) is increasingly used in medical imaging for disease diagnosis.
• The deployment of AI in clinical settings is hindered by its vulnerability to adversarial attacks.

Purpose of the Study:

• To systematically evaluate the susceptibility of the VolBrain neuroimaging diagnostic platform to universal black-box adversarial attacks.
• To assess the effectiveness of different adversarial attack methods, including FGSM and DeepFool, in degrading VolBrain's diagnostic performance.

Main Methods:

• Generated adversarial perturbations using a surrogate convolutional neural network (CNN) trained on a different dataset and architecture.
• Employed Fast Gradient Sign Method (FGSM) and DeepFool attacks to create universal black-box perturbations.
• Evaluated the impact of these perturbations on VolBrain's diagnostic performance.

Main Results:

• Adversarial perturbations reliably degraded VolBrain's diagnostic performance.
• DeepFool-based attacks were particularly effective for comparable perturbation sizes.
• A Mean Attack approach also demonstrated effectiveness, indicating vulnerability to universal attacks.

Conclusions:

• VolBrain is susceptible to universal black-box adversarial attacks, even without internal system knowledge.
• These findings underscore the significant risks posed by such attacks in medical AI.
• There is an urgent need for robust defense mechanisms and further research into the adversarial robustness of medical AI systems.