AI Security for SaMD: Fortifying AI Systems against Attacks in Medical Technologies

AI algorithms empower healthcare and medical technologies in developing new drugs, interpreting medical images, cleaning up electronic patient charts and in transforming radiology, retinopathy, high-performance medicine and much more. However, the ability for AI/ML software are uniquely susceptible to advanced techniques to subvert otherwise-reliable machine-learning systems—so called adversarial attacks or AI attacks. The questions AI SaMD face are the following: Do you have a way to detect AI attacks (including intentional failures) and act upon them? Could you demonstrate preventive measures and reasonable assurance against attacks? Will you be able to alert relevant constituents when such an attack happens? These challenges are compounded in the medical context with a potential for medical ML systems to compromise patient safety and healthcare accessibility.