Pretraining Data Exposure in Large Language Models: A Survey of Membership Inference, Data Contamination, and Security Implications

Abstract

Large Language Models (LLMs) have become the predominant paradigm in NLP, advancing both research and industry. As model sizes and pretraining data grow, concerns about Pretraining Data Exposure (PDE) increase due to the scale and opacity of training datasets. PDE refers to determining whether specific data appeared in an LLM's pretraining corpus. It is critical for ensuring evaluation integrity and protecting privacy, intersecting two key areas: data contamination and membership inference. Though conceptually related, these areas have often been studied in isolation. This paper offers the first unified survey of both under the PDE framework. We formalize PDE across exposure levels, review attack and defense methods, synthesize empirical findings, and highlight open challenges and future research directions.

accepted by NLDB 2025

References in corpus (45)

• Quantifying Memorization Across Neural Language Models
• LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code
• Copyright Traps for Large Language Models
• LiveBench: A Challenging, Contamination-Limited LLM Benchmark
• Leak, Cheat, Repeat: Data Contamination and Evaluation Malpractices in Closed-Source LLMs
• Don't Make Your LLM an Evaluation Benchmark Cheater
• Detecting Pretraining Data from Large Language Models
• Do Membership Inference Attacks Work on Large Language Models?
• Practical Membership Inference Attacks against Fine-tuned Large Language Models via Self-prompt Calibration
• Quantifying Privacy Risks of Masked Language Models Using Membership Inference Attacks
• NLP Evaluation in trouble: On the Need to Measure LLM Data Contamination for each Benchmark
• Investigating Data Contamination in Modern Benchmarks for Large Language Models
• A Survey on Data Contamination for Large Language Models
• Task Contamination: Language Models May Not Be Few-Shot Anymore
• Trained Without My Consent: Detecting Code Inclusion In Language Models Trained on Code
• Proving Test Set Contamination in Black Box Language Models
• Stop Uploading Test Data in Plain Text: Practical Strategies for Mitigating Data Contamination by Evaluation Benchmarks
• A Comprehensive Survey of Contamination Detection Methods in Large Language Models
• Watermarking Text Data on Large Language Models for Dataset Copyright
• An Open Source Data Contamination Report for Large Language Models
• Decoding Secret Memorization in Code LLMs Through Token-Level Characterization
• Concerned with Data Contamination? Assessing Countermeasures in Code Language Model
• Generalization or Memorization: Data Contamination and Trustworthy Evaluation for Large Language Models
• ConStat: Performance-Based Contamination Detection in Large Language Models
• Evading Data Contamination Detection for Language Models is (too) Easy
• Code Membership Inference for Detecting Unauthorized Data Use in Code Pre-trained Language Models
• TRUCE: Private Benchmarking to Prevent Contamination and Improve Comparative Evaluation of LLMs
• Sampling-based Pseudo-Likelihood for Membership Inference Attacks
• Can Watermarking Large Language Models Prevent Copyrighted Text Generation and Hide Training Data?
• Towards Data Contamination Detection for Modern Large Language Models: Limitations, Inconsistencies, and Oracle Challenges
• VarBench: Robust Language Model Benchmarking Through Dynamic Variable Perturbation
• AdEval: Alignment-based Dynamic Evaluation to Mitigate Data Contamination in Large Language Models
• A Taxonomy for Data Contamination in Large Language Models
• Detecting Training Data of Large Language Models via Expectation Maximization
• Order of Magnitude Speedups for LLM Membership Inference
• LatestEval: Addressing Data Contamination in Language Model Evaluation through Dynamic and Time-Sensitive Test Construction
• C${}^2$LEVA: Toward Comprehensive and Contamination-Free Language Model Evaluation
• MemHunter: Automated and Verifiable Memorization Detection at Dataset-scale in LLMs
• Underestimated Privacy Risks for Minority Populations in Large Language Model Unlearning
• Privacy Backdoors: Enhancing Membership Inference through Poisoning Pre-trained Models
• Does Your Neural Code Completion Model Use My Code? A Membership Inference Approach
• AntiLeakBench: Preventing Data Contamination by Automatically Constructing Benchmarks with Updated Real-World Knowledge
• The Mosaic Memory of Large Language Models
• RWKU: Benchmarking Real-World Knowledge Unlearning for Large Language Models
• Inference-Time Decontamination: Reusing Leaked Benchmarks for Large Language Model Evaluation