Differentially Private Training in the LLM Era

Overview

Differentially private stochastic gradient descent (DP-SGD) has been extensively studied in recent years from a variety of perspectives. However, applying DP-SGD to large language model (LLM) training introduces several additional challenges. First, unlike the end-to-end training pipeline commonly used in classical deep learning, LLM development typically involves multiple training stages, each of which may impose different privacy requirements. Second, DP-SGD is difficult to scale to large models due to its inherent computational and memory overhead.

In this talk, I will present our recent work on addressing these challenges from both theoretical and systems perspectives. I will first discuss how to make DP-SGD scalable for LLM pretraining with little to no sacrifice in throughput or memory efficiency. I will then show how to enable scalable differentially private fine-tuning for extremely large LLMs under limited computational resources. Next, I will present theoretical limitations of differentially private alignment, and show how these barriers can be addressed through new structure-aware methods, together with corresponding lower-bound analyses. Finally, I will discuss future research directions along these lines.