Lecture 4: Training Data Poisoning

In this lecture, starting with two real-world examples of training data poisoning attacks, first we define data poisoning attacks; differentiate indiscriminate poisoning and backdoor poisoning attacks; go over threat model considerations and poisoning alternatives. We then cover poisoning defense approaches with focus on training data sanitization, poisoning/backdoor detection, and robust training approaches to defend poisoning. Course website: https://trustworthy-ml-course.github.io/