Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Learn advanced techniques for privacy auditing of Large Language Models in this Google TechTalk presented by postdoctoral researcher Ashwinee Panda. Discover why current privacy auditing methods for LLMs are insufficient due to their reliance on basic canary generation approaches, which result in weak membership inference attacks and potentially cause researchers to underestimate privacy leakage from training models on sensitive datasets. Explore novel methods for generating effective canaries across various threat models and examine experimental results across multiple LLM families that demonstrate how these new approaches establish a higher standard for detecting privacy leakage. Gain insights into the critical intersection of machine learning privacy and large language model security from a researcher who completed his PhD at Princeton under Prof. Prateek Mittal and currently works as a postdoctoral fellow with Tom Goldstein at UMD College Park.
