CS 91R : Security in AI Systems — Fall 2026

By the end of this seminar, students should be able to:

There is no single required textbook. All primary readings are research papers linked in the schedule.

This course is a discussion-driven seminar. Tuesday lectures cover core technical concepts and common pitfalls in the readings.

Each student will present four papers during the semester, approximately one to two per section. Presentations are 15 minutes plus 10 minutes of Q&A. Attendance and active participation in discussion is expected in both class and lab; more than two unexcused absences will affect your participation grade.

Wednesday lab periods will host quizzes on alternate weeks, hands-on lab work, and some lecture material.

If you cannot attend class or lab, or cannot submit an assignment on time, email me before the deadline whenever possible. Each student gets one no-questions-asked 48-hour extension per semester on a weekly lab (not on presentations, quizzes, the two-week lab, or the final project). Beyond that, extensions are granted only for documented illness, family emergencies, or officially recognized religious observance.

Many techniques in this course — adversarial examples, model extraction, fuzzing, deepfake generation — can cause real harm if used against systems or people without authorization. You may not run any technique from this course against any system, model, or dataset you do not own or have explicit written permission to test. All labs are designed to be done on the course infrastructure or on models and datasets we provide. If you are unsure whether an experiment is appropriate, ask before running it. Unauthorized testing of live services (including public LLM APIs) can violate the Computer Fraud and Abuse Act and terms of service — with legal, academic, and career consequences.

Academic honesty is required in all your work. Under no circumstances may you hand in work done with (or by) someone else under your own name. Your code should never be shared with anyone; you may not examine or use code belonging to someone else, nor may you let anyone else look at or make a copy of your code. This includes, but is not limited to, obtaining solutions from students who previously took the course or code that can be found online. You may not share solutions after the due date of the assignment or make them publicly available anywhere (e.g. public GitHub repository).

Discussing ideas and approaches to problems with others on a general level is fine (in fact, we encourage you to discuss general strategies with each other), but you should never read anyone else’s code or let anyone else read your code. All code and written homeworks you submit must be your own with the following permissible exceptions: code distributed in class, code found in the course text book, and code worked on with an assigned partner. In these cases, you should always include detailed comments that indicates on which parts of the assignment you received help, and what your sources were.

Use of generative AI in this course: Because this is a course about AI systems, you may use AI assistants (ChatGPT, Copilot, Claude, etc.) as a study aid — for explanation, for understanding, and for grammar checking. You must: (1) disclose which AI tools you used and how, in a short note at the top of each submission; (2) understand and be able to explain every line of code you submit; (3) never paste model output verbatim as your own writing.

You may not use AI to generate text for your paper presentations or for your final project reports. Undisclosed or unedited AI output will be treated as an academic-integrity violation.

Failure to abide by these rules constitutes academic dishonesty and will lead to a hearing of the College Judiciary Committee. According to the Faculty Handbook:

The spirit of this policy applies to all course work, including code, homework solutions (e.g., proofs, analysis, written reports), and exams. Please contact me if you have any questions about what is permissible in this course.

Students must strictly adhere to the following policy, which applies to all exams taken in a Computer Science course at Swarthmore:

Exam takers must place all non-essential items at the front of the room (or other designated area). Unless otherwise permitted, students may not have any electronic devices or course materials in their possession during the entirety of the exam. This includes cell phones, tablets, laptops, smart watches, course notes, articles and books, among others. These items should be placed at the front of the room near the proctor. If you need to leave the room during the exam, you must obtain permission from an instructor first. Any non-permitted discussion or aide in regards to exam material will result in immediate forfeiture of the exam and a report to the College Judiciary Committee. Please discuss any concerns or accommodations with your instructor prior to starting the exam.

If you believe you need accommodations for a disability or a chronic medical condition, please visit the Student Disability Services website for details about the accommodations process. Since accommodations require early planning and are not retroactive, contact Student Disability Services as soon as possible. You are also welcome to contact me, Vasanta, privately to discuss your academic needs. However, all disability-related accommodations must be arranged, in advance, through Student Disability Services.

To receive an accommodation for a course activity you must have an official Accommodations Letter and you need to meet with me to work out the details of your accommodation at least two weeks prior to any activity requiring accommodations.

You are also welcome to contact me, privately to discuss your academic needs. However, all disability-related accommodations must be arranged, in advance, through Student Disability Services.