Please join us for a UW Data Science Seminar on Tuesday, November 14th from 4:30 to 5:20 p.m. PST. The seminar will feature Niloofar Mireshghallah, a Postdoctoral Scholar with the UW Allen Center for Computer Science & Engineering.
This event will take place in the Physics/Astronomy Auditorium 102 (PAA A102) on the University of Washington campus.
“Privacy Auditing and Protection in Large Language Models”
Abstract: Large language Models (LLMs, e.g., GPT-3, OPT, TNLG,…) are shown to have a remarkably high performance on standard benchmarks, due to their high parameter count, extremely large training datasets, and significant compute. Although the high parameter count in these models leads to more expressiveness, it can also lead to higher memorization, which, coupled with large unvetted, web-scraped datasets can cause different negative societal and ethical impacts such as leakage of private, sensitive information and generation of harmful text. In this talk, we will go over how these issues affect the trustworthiness of LLMs, and zoom in on how we can measure the leakage and memorization of these models, and mitigate it through differentially private training. Finally we will discuss what it would actually mean for LLMs to be privacy preserving, and what are the future research directions on making large models trustworthy.
Biography: Niloofar Mireshghallah is a post-doctoral scholar at the Paul G. Allen Center for Computer Science & Engineering at University of Washington. She received her Ph.D. from the CSE department of UC San Diego in 2023. Her research interests are Trustworthy Machine Learning and Natural Language Processing. She is a recipient of the National Center for Women & IT (NCWIT) Collegiate award in 2020 for her work on privacy-preserving inference, a finalist of the Qualcomm Innovation Fellowship in 2021 and a recipient of the 2022 Rising star in Adversarial ML award.
The UW Data Science Seminar is an annual lecture series at the University of Washington that hosts scholars working across applied areas of data science, such as the sciences, engineering, humanities and arts along with methodological areas in data science, such as computer science, applied math and statistics. Our presenters come from all domain fields and include occasional external speakers from regional partners, governmental agencies and industry.
The 2022-2023 seminars will be held in person, and are free and open to the public.