The Surprising Effectiveness of Membership Inference with Simple N-Gram Coverage

This seminar presentation from April 24, 2025, features Skyler Hallinan from USC discussing a novel approach to membership inference attacks on language models. Learn about the N-Gram Coverage Attack, a method that requires only text outputs from target models, enabling attacks on completely black-box systems like GPT-4. Discover how this technique leverages the observation that models tend to memorize and generate text patterns commonly found in their training data. The presentation demonstrates how this approach outperforms other black-box methods and achieves comparable or better results than state-of-the-art white-box attacks despite limited access. Explore findings showing that attack success rates scale with compute budget and that newer models like GPT-4o exhibit increased robustness to membership inference, suggesting improved privacy protections. Hallinan, a Ph.D. student at USC advised by Xiang Ren, focuses on building trustworthy AI systems through data-centric approaches, with experience as a research intern at Apple and Amazon.