Guest lecture by Dr. Franziska Boenisch from CISPA Helmholtz Center for Information: Understanding and Mitigating Memorization in Foundation Models.
Date and time: June 25, 2026 from 13:15-14:00 EEST
Event type: Hybrid
Venue: Physicum, room D112, University of Helsinki Kumpula Campus, Pietari Kalmin katu 5
ZOOM: LINK, Passcode: 537927
Abstract: Memorization occurs when machine learning models store and reproduce specific training examples at inference time, raising serious concerns for privacy and intellectual property protection. In this talk, I will explore methods to measure memorization in foundation models that go beyond coarse-grained binary scores. Using these proposed metrics, I will provide insights into memorization across self-supervised vision encoders, CLIP-style multi-modal encoders, and state-of-the-art diffusion models. Beyond measurement, I will demonstrate how to identify the specific neurons responsible for triggering memorized outputs. By selectively pruning these identified neurons, we can prevent verbatim reproduction of training data without significantly degrading the model's generation capabilities. Finally, building on these findings, I will address a fundamental question: Is memorization a localized phenomenon within foundation models' parameters? I will present our initial evidence suggesting it is not. Overall, my research contributes to understanding the mechanisms behind memorization in foundation models and developing practical strategies to mitigate unintended memorization.
Speaker bio: Franziska Boenisch is a tenure-track faculty at the CISPA Helmholtz Center for Information Security, where she co-leads the SprintML lab. Her research focuses on private and trustworthy machine learning; during her Ph.D. at Freie Universität Berlin and Fraunhofer AISEC she pioneered the notion of individualized privacy for ML. Before joining CISPA, she was a Postdoctoral Fellow at the University of Toronto and the Vector Institute. She received an ERC Starting Grant in 2025 for research on privacy in foundation models and has been recognized with the Fraunhofer ICT Dissertation Award (2023), a GI Junior Fellowship (2024), and a Werner‑von‑Siemens Fellowship (2025).