Diagnostic AI may be vulnerable to cyberattacks, UPMC study shows

Researchers from the University of Pittsburgh published a study in Nature Communications that explored the ability of adversarial images to deceive an AI model designed to diagnose breast cancer. The researchers were able to create fake mammogram images that fooled both the AI model and human radiologists, highlighting the potential patient safety risks of such attacks. The study emphasizes the need to develop safer and more robust AI models in medical contexts.