PrivacyMind: Large Language Models Can Be Contextual Privacy Protection Learners

TL;DR

PrivacyMind introduces CPPLM, a paradigm for fine-tuning LLMs that injects domain knowledge while protecting contextual PII during inference. The approach combines corpus curation, penalty-based unlikelihood, a PII classifier, instruction-based tuning, and DPO, underpinned by a theoretical analysis of information loss in corpus curation. Empirical results across biomedical and clinical datasets show instruction-based tuning with positive and negative examples achieving favorable privacy-utility Pareto frontiers, suggesting LLMs can learn domain knowledge without leaking sensitive data. The work highlights practical applicability and provides code and data to enable implementation in privacy-sensitive settings.

Abstract

The proliferation of Large Language Models (LLMs) has driven considerable interest in fine-tuning them with domain-specific data to create specialized language models. Nevertheless, such domain-specific fine-tuning data often contains contextually sensitive personally identifiable information (PII). Direct fine-tuning of LLMs on this data without privacy protection poses a risk of data leakage of sensitive PII during inference time. To address this challenge, we introduce Contextual Privacy Protection Language Models (PrivacyMind), a novel paradigm for fine-tuning LLMs that effectively injects domain-specific knowledge while safeguarding inference-time data privacy. Our work offers a theoretical analysis for model design and benchmarks various techniques such as corpus curation, penalty-based unlikelihood in training loss, instruction-based tuning, etc. Extensive experiments across diverse datasets and scenarios demonstrate the effectiveness of our approaches. In particular, instruction tuning with both positive and negative examples stands out as a promising method, effectively protecting private data while enhancing the model's knowledge. Our work underscores the potential for Large Language Models as robust contextual privacy protection learners. The complete code and data for the work can be found at https://github.com/Yijia-Xiao/PrivacyMind.

Figures (11)

• Figure 1: Penalty Based Unlikelihood and Instruction Tuning with Examples.
• Figure 2: Pareto Frontier.
• Figure 3: Vanilla, Removal, Substitution.
• Figure 4: Pareto Frontier on medical_flashcards, LLaMA2-7B
• Figure 5: Pareto Frontier on medical_flashcards, LLaMA2-13B