From the course: Security Risks in AI and Machine Learning: Categorizing Attacks and Failure Modes

Unlock this course with a free trial

Join today to access over 24,900 courses taught by industry experts.

Reprogramming

Reprogramming

From the course: Security Risks in AI and Machine Learning: Categorizing Attacks and Failure Modes

Start my 1-month free trial Buy for my team

Reprogramming

“

- [Instructor] Another way to attack AI models is to use perturbation to reprogram the system to perform a task it wasn't intended to. Note that this is not to be confused with transfer learning, which is non-adversarial, and refers to when an ML-based AI system transfers knowledge to a different problem space. For example, when vision models trained on general photos like cats, dogs and cars, are adapted to detect disease from medical images. When attackers attempt to reprogram AI, they send the models queries, inducing them to solve new and unintended tasks. For example, consider CAPCTHA. CAPTCHAs often require users to identify 2D images like crosswalks, traffic signals, or a mountain, but if an image classifier can accurately identify those images as well, or even better than a human, someone could use that to give a bot access as though it were a human. This kind of attack could make CAPTCHA useless in instances…

Contents