Privacy-Preserving Low-Rank Instruction Tuning for Large Language Models via DP-LoRA

Abstract

This paper proposes DP-LoRA, an instruction tuning algorithm that combines differential privacy with low-rank adaptation to address the challenges of privacy risks and performance retention in large-scale language models for instruction-following tasks. The method embeds low-rank adaptation modules on top of a frozen pretrained backbone and integrates differential privacy through gradient clipping and noise injection to strictly control the privacy budget while ensuring effective model updates. A systematic analysis is conducted from three perspectives: hyperparameter sensitivity, environmental sensitivity, and data sensitivity.   The study examines the impact of privacy budgets on various aspects, including perplexity, membership inference attack success rates, and instruction adherence. It also investigates the performance changes during communication rounds and bandwidth constraints. Additionally, the study explores the effects of instruction diversity and task mixture on privacy consumption and performance. Experimental results show that DP-LoRA reduces perplexity, improves instruction adherence, and mitigates privacy risks while maintaining robustness under distributed and multi-task conditions. This research not only achieves a unified balance between privacy protection and performance but also demonstrates strong adaptability in multidimensional sensitivity experiments, providing systematic validation and empirical evidence for the application of differential privacy in instruction tuning for large models.