Adversarial Model Analysis / Analiza adwersaryjna, bezpieczeństwo i wyjaśnialność systemów sztucznej inteligencji - 2024/25

Adversarial Model Analysis course for Machine Learning (MSc) specialisation at the University of Warsaw.

Spring semester 2024/25 @pbiecek @sobieskibj

Meetings

Plan for the spring semester 2024/2025. MIM_UW classes are on Fridays.

• 2025-02-28 -- these classes won't happen! we will meet next week!
• 2025-03-07 -- Introduction
• 2025-03-14 -- Security (NIST, OWASP, MITTRE + ESA)
• 2025-03-21 -- Security (Snowflake)
• 2025-03-28 -- Safety (LLMs) Wiktoria o perswazyjności modeli
• 2025-04-04 -- Safety (LLMs)
• 2025-04-11 -- Adversarial attacks on models and explanations (Hubert based on the survey)
• 2025-04-25 -- (ICLR)
• 2025-05-07 -- PROJECT: first presentation (prerecorded videos)
• 2025-05-16 -- Adversarial analysis of prototypical models (Hubert based on "Birds look like cars")
• 2025-05-23
• 2025-05-30 -- Student presentations of research papers
• 2025-06-06
• 2025-06-13 -- PROJECT: final presentation (in-person presentations)

How to get a good grade

The final grade is based on activity in four areas:

• Project 60% (first part: 0-20, 0-40 final part)
• Exam 15% (0-15)
• Presentation 15% (0-15)
• Activity 15% (0-15) - three small tasks (security, safety, adversarial), 5 points each

In total you can get from 0 to 100 points. 51 points are needed to pass this course.

Grades:

• 51-60: (3) dst
• 61-70: (3.5) dst+
• 71-80: (4) db
• 81-90: (4.5) db+
• 91-100: (5) bdb

Project

Pick a model for adversarial analysis

• Bielik
• CLIP
• Stable Diffusion

Presentation (0-15 points)

Choose an article from the last two years published at the A* conference or another equally interesting source. Example papers

• list of papers will be provided later

References

• Adversarial Model Analysis https://ama.drwhy.ai/
• Adversarial Machine Learning https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-2e2023.pdf
• AI Security Institute https://www.aisi.gov.uk/