•   When: Friday, November 14, 2025 from 11:00 AM to 12:00 PM
•   Speakers: Vitaly Shmatikov
•   Location: JC George's
•   Export to iCal

ABSTRACT: Embeddings are vector representations of inputs into machine learning models.  In this talk, I will describe how these vectors encode meaning within ML systems, how they can be inverted into original inputs, and the latest results on learning and translating the universal geometry of embeddings.  Our vec2vec method translates embeddings from one vector space to another without any paired data, encoders, or predefined sets of potential matches.  This is achieved by learning a universal semantic representation, which provides empirical evidence for the so-called Platonic Representation Hypothesis and makes black-box embedding inversion possible. I will discuss the implications of embedding inversion for the security of vector databases and AI systems in general.
This is joint work with Jack Morris, Rishi Jha, Collin Zhang, and others.

BIO: Vitaly Shmatikov is a professor of computer science at Cornell Tech. Before Cornell, he worked at the University of Texas and SRI International. His research areas include digital privacy, computer security, and security and privacy issues in machine learning.  Research by Shmatikov, his students, and collaborators received the Caspar Bowden PET Award for Outstanding Research in Privacy Enhancing Technologies three times; Test-of-Time Awards from the IEEE Symposium on Security and Privacy, the ACM Conference on Computer and Communications Security (twice), and the ACM/IEEE Symposium on Logic in Computer Science; as well as several outstanding and distinguished paper awards, most recently from USENIX Security (twice) and EMNLP.