Preserving Privacy in Large Language Models: A Survey on Current Threats and Solutions

2024-08-10

arXiv.org

AI Summary

This work systematically uncovers privacy risks across the full lifecycle of large language models (LLMs), encompassing twelve attack types—including training data memorization leakage, membership inference, and attribute inference. Method: It introduces the first end-to-end classification framework for LLM privacy risks and mitigations, spanning data preprocessing, training, inference, and model decommissioning. The framework integrates diverse technical approaches—differential privacy, machine unlearning, data sanitization and synthesis, adversarial modeling, and privacy evaluation—while rigorously characterizing their applicability boundaries and synergistic interactions. A utility–privacy–overhead trade-off principle is proposed to guide practical deployment. Contribution/Results: The study synthesizes over 30 defense techniques into an actionable, implementation-oriented privacy-enhancement roadmap; its findings have been incorporated into multiple industry LLM security guidelines.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Security and Privacy: Security and privacy of machine learning and AI applications

Abstract

Large Language Models (LLMs) represent a significant advancement in artificial intelligence, finding applications across various domains. However, their reliance on massive internet-sourced datasets for training brings notable privacy issues, which are exacerbated in critical domains (e.g., healthcare). Moreover, certain application-specific scenarios may require fine-tuning these models on private data. This survey critically examines the privacy threats associated with LLMs, emphasizing the potential for these models to memorize and inadvertently reveal sensitive information. We explore current threats by reviewing privacy attacks on LLMs and propose comprehensive solutions for integrating privacy mechanisms throughout the entire learning pipeline. These solutions range from anonymizing training datasets to implementing differential privacy during training or inference and machine unlearning after training. Our comprehensive review of existing literature highlights ongoing challenges, available tools, and future directions for preserving privacy in LLMs. This work aims to guide the development of more secure and trustworthy AI systems by providing a thorough understanding of privacy preservation methods and their effectiveness in mitigating risks.

Problem

• Privacy threats in Large Language Models
• Solutions for privacy preservation in LLMs
• Challenges in secure AI system development

Innovation

• Anonymizing training datasets for privacy
• Implementing differential privacy in training
• Machine unlearning post-training for security