Important communication will be via Canvas.

Important Note to Interested Students: The current class syllabus and grading are not yet finalized. Depending on the class size and the time required for paper reading and discussion, we might have to adjust the lecture part of the course.

Course Description

This course is designed to provide a comprehensive introduction to machine learning (ML) security and privacy research topics for graduate students. Students will learn how to formulate many of the current issues in ML in a mathematical way, as well as design algorithmic strategies to mitigate security and privacy vulnerabilities in ML. As a result, the course is half lecture-style, half seminar-style. Lectures are intended to lay down the necessary tools (e.g., hypothesis testing, formal verification, causality, etc.) for reading more advanced CS papers on presented topics (which is the seminar part of the course). The course aims to strike a balance between mathematical rigor and practical implementation and relevance. Hence, there will be coding assignments and a group project (which may involve implementation as well).

Students are expected to engage with the material being taught, read suggested research papers about the presented topics, and actively engage in discussions during class. Readings will mostly consist of CS papers.

(Good to have) Prerequisites: Probability Theory, Introduction to Machine Learning (or equivalent basic AI or ML class).

Intended Learning Outcomes

By the end of this class, students will be able to:

• Understand the use of optimization, probability & statistics, causality and algebra in the field of security and privacy of machine learning systems.
• Be familiar with one of the ML frameworks (e.g., Pytorch, Tensorflow) to train and evaluate the security and privacy of ML.
• Compare threat models and trust assumptions of security and privacy attacks and defenses.
• Explain challenges of safely integrating AI in systems and illustrate with examples.

Assessments

In-class quizzes

• These quizzes will be announced beforehand and taken at the start of the class online, on Canvas. Please do not be late and please bring your laptop or tablet to take it online. If you cannot do that for any reason, please email the TAs 2-3 days before the class. Quizzes consist of questions about the material that is being taught during that week’s topic and previous lecture topics. These are intended to allow students to consolidate their understanding of the material being taught, and to assess the effectiveness of teaching.
• There will be conceptual quizzes throughout the semester on GradeScope (tentative).