Cohere For AI - Guest Speaker: Anshuman Suri, Ph.D. Candidate 

Join the NLP group as they welcome Anshuman Suri to present "Do Membership Inference Attacks Work on Large Language Models?" (https://arxiv.org/abs/2402.07841)

Abstract:
Membership inference attacks (MIAs) aim to discern whether a specific data point belongs to the training data of a target model. While extensive research has addressed MIAs in traditional machine learning models, limited attention has been paid to their study in the pre-training data of large language models (LLMs). In this presentation, I will discuss our investigation into MIAs across a range of LLMs trained on the Pile dataset. Our findings indicate that MIAs only marginally outperform random guessing across various model sizes and domains. I will then delve into how the observed poor performance can be attributed to the unique characteristics of LLM training and the interpretation of membership within text-based generative models.

Add event to calendar

Apple Google Office 365 Outlook Outlook.com Yahoo