The Math of ML Security

Sven Cattell (Elastic)

02-Jun-2021, 19:00-20:00 (3 years ago)

Abstract: Security is a cat and mouse game. Attackers innovate to bypass defenders, and defenders innovate to catch the new attacks. One of the promises of AI is its ability to adapt for us. However, the machine learning models we deploy are trained on a chronological snapshot of the ever-changing data. They memorize and generalize well on that snapshot but are unreliable when the landscape shifts, or have various adversarial examples and other holes that let attackers to bypass the ML models. This talk will focus on how this plays out mathematically on the large datasets we use to create these models. We will also talk about how I transitioned into this industrial space, from a PhD in equivariant algebraic topology and advice that might make things easier for future grad students looking towards ML.

Bio: Sven Cattell is a Senior Security Data Scientist at Elastic. He received his mathematics PhD from Johns Hopkins University where his thesis focused on equivariant algebraic topology. During his post doctoral his focused shifted to the geometry of machine learning. While working on his postdoc he co-founded the AI Village at DEFCON which will be at DEFCON for the fourth time this year. He also built a math exhibit for the National Science Fair in DC and a game to teach kids about disinformation and spam. He now works at Elastic Security on their malware models trying to improve the model and secure it against adversarial attacks.

Mathematics

Audience: advanced learners


Graduate Online Seminar Series (GOSS)

Series comments: Meeting Password: MATHGOSS

Announcement mailing list: groups.google.com/g/goss2021

Website: dzackgarza.com/GOSS/2021/

Recordings: www.youtube.com/watch?v=n3xhHlOzFPM&list=PLkscP0p2V2U5J-Gc4foDjQxVD-4c0dVU4

Organizer: D. Zack Garza*
*contact for this listing

Export talk to