Deep learning models are increasingly used in radiology to detect changes in X-rays, CTs, and MRIs. However, these systems are highly vulnerable to adversarial attacks—small, crafted perturbations that can mislead models while remaining imperceptible to human observers.

The new preprint focuses on a specific class of these threats: directional adversarial attacks. The authors propose a method for robust detection of such attacks, aiming to make radiological AI systems more reliable in clinical settings.

The abstract does not provide details on the detection technique itself, but it underscores the importance of defending medical imaging models against malicious input. The work is available on arXiv under the identifier 2610.15628.