Revisiting Privacy, Utility, and Efficiency Trade-offs when Fine-Tuning Large Language Models

TL;DR

This paper explores the trade-offs between privacy, utility, and efficiency in fine-tuning large language models, revealing that efficient methods like LoRA can mitigate privacy risks similarly to private training methods, challenging previous assumptions.

Contribution

It demonstrates empirically that efficient fine-tuning methods can reduce privacy risks comparable to differential privacy, contradicting the belief that privacy and efficiency are at odds.

Findings

Efficient fine-tuning methods like LoRA mitigate privacy risks similar to differential privacy.

Privacy and efficiency objectives are not necessarily conflicting during fine-tuning.

Extensive evaluations across multiple models and datasets support these conclusions.

Abstract

We study the inherent trade-offs in minimizing privacy risks and maximizing utility, while maintaining high computational efficiency, when fine-tuning large language models (LLMs). A number of recent works in privacy research have attempted to mitigate privacy risks posed by memorizing fine-tuning data by using differentially private training methods (e.g., DP), albeit at a significantly higher computational cost (inefficiency). In parallel, several works in systems research have focussed on developing (parameter) efficient fine-tuning methods (e.g., LoRA), but few works, if any, investigated whether such efficient methods enhance or diminish privacy risks. In this paper, we investigate this gap and arrive at a surprising conclusion: efficient fine-tuning methods like LoRA mitigate privacy risks similar to private fine-tuning methods like DP. Our empirical finding directly contradicts prevailing wisdom that privacy and efficiency objectives are at odds during fine-tuning. Our finding is established by (a) carefully defining measures of privacy and utility that distinguish between memorizing sensitive and non-sensitive tokens in training and test datasets used in fine-tuning and (b) extensive evaluations using multiple open-source language models from Pythia, Gemma, Llama, and Qwen families and different domain-specific datasets.