Context-Aware Membership Inference Attacks against Pre-trained Large Language Models

2024-09-11

arXiv.org

AI Summary

Membership inference attacks (MIAs) against pre-trained large language models (LLMs) face a fundamental challenge: existing classification-based approaches ignore the generative nature of LLMs and fail to capture their sequence-level memorization behavior. Method: We propose a context-aware MIA framework that, for the first time, applies statistical hypothesis testing to analyze token-level perplexity evolution trajectories—leveraging subsequence perplexity modeling, context-sensitive memory representation, and contrastive loss dynamics analysis. Contribution/Results: Our method breaks from conventional transfer-based attack paradigms and achieves significant improvements over state-of-the-art loss-based MIAs across multiple mainstream pre-trained LLMs, markedly enhancing attack accuracy. It reveals strong context-dependent heterogeneity in pre-training data memorization, offering novel insights into LLM memorization mechanisms and associated privacy risks, along with an effective analytical tool for rigorous privacy evaluation.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Search and Optimization: Learning to Search

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Search and Retrieval-Augmented AI: Large language models for search
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Prior Membership Inference Attacks (MIAs) on pre-trained Large Language Models (LLMs), adapted from classification model attacks, fail due to ignoring the generative process of LLMs across token sequences. In this paper, we present a novel attack that adapts MIA statistical tests to the perplexity dynamics of subsequences within a data point. Our method significantly outperforms prior loss-based approaches, revealing context-dependent memorization patterns in pre-trained LLMs.

Problem

• Membership inference attacks on LLMs' training data
• Adapting statistical tests to perplexity dynamics
• Revealing context-dependent memorization patterns

Innovation

• Context-aware MIA using perplexity dynamics
• Adapts statistical tests to subsequences
• Reveals context-dependent memorization patterns