Reproduce “Privacy-preserving Fine-tuning of Large Language Models through Flatness” on OpenResearch

About this paper

Tiejin Chen, Longchao Da, Huixue Zhou, Pingzhi Li, Kaixiong Zhou, Tianlong Chen, Hua Wei

The paper introduces DP-Flat, a framework for privacy-preserving fine-tuning of large language models that enhances generalization by promoting flatness in the loss landscape. It addresses the privacy-utility trade-off in differentially private training, demonstrating substantial performance improvements in both white-box and black-box settings while maintaining strong privacy guarantees.

Key insights

• Differential Privacy (DP) training for Large Language Models (LLMs) tends to converge to sharper minima in the loss landscape, leading to reduced generalization performance.
• Actively promoting flatness in the loss landscape during DP fine-tuning can effectively mitigate the privacy-utility trade-off, improving model generalization without compromising privacy.
• Privacy-preserving fine-tuning can be extended to black-box LLMs by leveraging zeroth-order optimization and cross-model knowledge distillation, addressing a critical practical limitation.

Summary by alphaXiv.