Select a calendar:
Filter April Events by Event Type:
SUNMONTUEWEDTHUFRISAT
Events for April 24, 2025
-
NL Seminar: The Surprising Effectiveness of Membership Inference with Simple N-Gram Coverage
Thu, Apr 24, 2025 @ 11:00 AM - 12:00 PM
Information Sciences Institute
Conferences, Lectures, & Seminars
Speaker: Skyler Hallinan, USC
Talk Title: The Surprising Effectiveness of Membership Inference with Simple N-Gram Coverage
Series: NL Seminar
Abstract: Meeting hosts only admit on-line guests that they know to the Zoom meeting. Hence, you’re highly encouraged to use your USC account to sign into Zoom. If you’re an outside visitor, please inform us at ((nlg-seminar-host(at)isi.edu) to make us aware of your attendance so we can admit you. Specify if you will attend remotely or in person at least one business day prior to the event. Provide your: full name, job title and professional affiliation and arrive at least 10 minutes before the seminar begins. If you do not have access to the 6th Floor for in-person attendance, please check in at the 10th floor main reception desk to register as a visitor and someone will escort you to the conference room location. Join Via Zoom: https://usc.zoom.us/j/96791099940?pwd=6kov3zTLAnD4JU49d1VtX4XNAZMcvs.1Meeting ID: 967 9109 9940Passcode: 840282
Membership inference attacks serves as useful tool for fair use of language models, such as detecting potential copyright infringement and auditing data leakage. However, many current state-of-the-art attacks require access to models' hidden states or probability distribution, which prevents investigation into more widely-used, API-access only models like GPT-4. In this work, we introduce N-Gram Coverage Attack, a membership inference attack that relies solely on text outputs from the target model, enabling attacks on completely black-box models. We leverage the observation that models are more likely to memorize and subsequently generate text patterns that were commonly observed in their training data. Specifically, to make a prediction on a candidate member, N-Gram Coverage Attack first obtains multiple model generations conditioned on a prefix of the candidate. It then uses n-gram overlap metrics to compute and aggregate the similarities of these outputs with the ground truth suffix; high similarities indicate likely membership. We first demonstrate on a diverse set of existing benchmarks that N-Gram Coverage Attack outperforms other black-box methods while also impressively achieving comparable or even better performance to state-of-the-art white-box attacks --- despite having access to only text outputs. Interestingly, we find that the success rate of our method scales with the attack compute budget --- as we increase the number of sequences generated from the target model conditioned on the prefix, attack performance tends to improve. Having verified the accuracy of our method, we use it to investigate previously unstudied closed OpenAI models on multiple domains. We find that more recent models, such as GPT-4o, exhibit increased robustness to membership inference, suggesting an evolving trend toward improved privacy protections.
Biography: Skyler Hallinan is a Ph.D. student in Computer Science at the University of Southern California where he is advised by Xiang Ren. His research aims to build trustworthy AI systems with robust reasoning capabilities via data-centric approaches. His work spans three core areas: understanding how data impacts downstream model behavior, safeguarding user data and privacy, and advancing model capabilities with better data. Previously, he was a research intern at Apple and Amazon, and received a B.S./M.S. in Computer Science from the University of Washington, where he was advised by Yejin Choi.
Host: Jonathan May and Katy Felkner
More Info: https://www.isi.edu/research-groups-nlg/nlg-seminars/
Webcast: https://usc.zoom.us/j/96791099940?pwd=6kov3zTLAnD4JU49d1VtX4XNAZMcvs.1Location: Information Science Institute (ISI) - Conf Rm#689
WebCast Link: https://usc.zoom.us/j/96791099940?pwd=6kov3zTLAnD4JU49d1VtX4XNAZMcvs.1
Audiences: Everyone Is Invited
Contact: Pete Zamar
Event Link: https://www.isi.edu/research-groups-nlg/nlg-seminars/
This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor. -
PhD Thesis Proposal - Neel Patel
Thu, Apr 24, 2025 @ 12:30 PM - 01:30 PM
Thomas Lord Department of Computer Science
University Calendar
Title: Combinatorial Optimization under Uncertainty, Incentives and Correlations.
Date and Time: 04/24, 12:30-1:30 pm
Location: GCS 502C
Committee Members: Shaddin Dughmi, David Kempe, Vatsal Sharan, Evi Micha and Greta Panova
Abstract:
This proposal considers algorithms for combinatorial problems, primarily those concerned with combinatorial selection under uncertainty and incentives, also known as stochastic selection problems. The core focus is on the two pivotal stochastic selection problems that include contention resolution schemes (CRS) and generalized prophet inequalities. Our contributions are twofold:
Our first contribution deepens the understanding of the stochastic selection problems beyond independent priors on the input and its implications on the famous matroid secretary conjecture. Our results completely characterize the CRS and prophet inequalities on matroids for pairwise independent priors. En route to proving our results, we develop techniques to sample exact pairwise independent vectors over a finite field from approximate pairwise independent vectors which later becomes a key ingredient for characterizing the difficult instance for binary matroid secretary conjecture.
The rest of the proposal aims to push the applications of the powerful algorithmic toolkit --- stochastic selection with a broader goal of identifying the algorithmic and economic questions that appear to be complex and algorithmically challenging, for which the techniques developed by online stochastic selection provide an alternative outlook, leading to more efficient and powerful algorithmic results. In this context, we prove the following key results:
1.) We obtain the first combinatorial generalized stationary prophet inequalities where our main result shows that the (offline) CRS plays a central role in the (online) stationary prophet inequality problem. This intriguing connection allows us to obtain several new algorithmic results as well as improves the existing results significantly.
2.) We systematically generalize the sparsification of stochastic matching problems to the general combinatorial structure. Here, we show that any combinatorial structure that exhibits `good’ CRS also exhibits strong stochastic sparsifiers.
3.) We obtain constant approximate delegation mechanisms for the principal-agent delegation problem with probing cost for a large class of combinatorial constraints. We obtain these mechanisms by reducing the delegation problem to the online version of CRS for combinatorial constraints.Location: Ginsburg Hall (GCS) - 502C
Audiences: Everyone Is Invited
Contact: Neel Patel
This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor. -
Semiconductors and Microelectronics Technology seminar - Wolfgang Maass, Thursday, April24th at 2pm in EEB 248
Thu, Apr 24, 2025 @ 02:00 PM - 03:30 PM
Ming Hsieh Department of Electrical and Computer Engineering
Conferences, Lectures, & Seminars
Speaker: Wolfgang Maas, Institute of Machine Learning and Neural Computation Graz University of Technology
Talk Title: Recent brain-data and theories suggest new ways of porting cognitive function into neuromorphic hardware through on-chip learning
Series: Semiconductors & Microelectronics Technology
Abstract: I will discuss experimental data and models for BTSP (Behavioral Time Scale Synaptic Plasticity), the only known mechanism for 1-shot learning in the brain. I also will explain how BTSP can be used to create content-addressable memories and to learn cognitive maps that enable flexible goal-directed behavior. References and a simple model for BTSP have already been published (Yujie Wu and Wolfgang Maass, Nat. Comm. 2025). The other material is unpublished.
Biography: Wolfgang Maass - Since 2023: Director of the ELLIS Unit Graz (ELLIS = European Lab for Learning and Intelligent Systems). 1992-2017 Founder and Head of the Institut fuer Grundlagen der Informationsverarbeitung (Institute of Theoretical Computer Science) at Graz University of Technology. Since 1991 Professor of Computer Science at the Graz University of Technology in Austria (since 2017 without teaching duties except education of Phd students, leader of research projects). 1986 - 1991 Professor of Computer Science at the University of Illinois in Chicago. 1984 - 1986 Associate Professor of Computer Science at the University of Illinois in Chicago. 1975 - 1984 Postdoc at the Ludwig-Maximilians-Universität in Munich, the Massachusetts Institute of Technology (MIT), the University of Chicago, and the University of California at Berkeley.
Host: Prof. Jayakanth Ravichandran, Prof. J. Joshua Yang, Prof. Chongwu Zhou, Prof. Stephen Cronin, and Prof. Wei Wu
More Information: Wolfgang Maass Flyer.pdf
Location: Hughes Aircraft Electrical Engineering Center (EEB) - 248
Audiences: Everyone Is Invited
Contact: Marilyn Poplawski
This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor.
