Challenges in Augmenting Large Language Models with Private Data

Explore the challenges of integrating private data into large language models in this 58-minute Google TechTalk presented by Ashwinee Panda from Princeton University. Delve into the concept of "neural phishing," a new data extraction attack that enables adversaries to target and extract personally identifiable information (PII) from models trained on user data. Learn about Differentially Private In-Context Learning, a framework proposed to coordinate independent LLM agents for answering user queries under differential privacy. Examine methods for obtaining consensus across potentially disagreeing LLM agents and investigate the privacy-utility tradeoff of different differential privacy mechanisms. Gain insights into the ongoing developments in LLM technology and their implications for both stronger adversaries and more robust systems.