Window-based Membership Inference Attacks Against Fine-tuned Large Language Models

Researchers developed WBC (Window-Based Comparison), a new membership inference attack method that significantly outperforms existing approaches by analyzing localized patterns in Large Language Models rather than global signals. The technique achieves 2-3 times better detection rates and exposes critical privacy vulnerabilities in fine-tuned LLMs through sliding window analysis and binary voting mechanisms.

• →WBC method outperforms traditional membership inference attacks by focusing on localized memorization patterns instead of global averaging.
• →The technique uses sliding windows with binary voting to identify training data with 2-3 times better detection rates.
• →Extensive testing across eleven datasets demonstrates superior AUC scores compared to established baselines.
• →Fine-tuned Large Language Models have significant privacy vulnerabilities that can be exploited through localized analysis.
• →The research challenges the current global-averaging paradigm in favor of context-specific membership detection.