Lecture 5: Membership Inference Attacks скачать в хорошем качестве

Lecture 5: Membership Inference Attacks

In this lecture, we focus on privacy risks in machine learning models with emphasis on membership inference attacks. We first highlight the increasing trend on the use of sensitive/private data for training machine learning models and enumerate privacy risks thereof. We then define membership inference and articulate why we should care and why it is possible. Most of the lecture is dedicated to attacks and defenses. On attacks, we cover two broad classes: shadow model-based and threshold-based attacks. On defenses, we study differentially private machine learning (the current 'gold' standard) and complementary approaches that rely on preemptive exclusion of member data points, and compare them via privacy/utility trade-offs. Course website: https://trustworthy-ml-course.github.io/