![[2502.18545] PII-Bench: Evaluating Query-Aware Privacy Protection Systems](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2502.18545] PII-Bench: Evaluating Query-Aware Privacy Protection Systems
Summary
The paper introduces PII-Bench, a novel framework for evaluating privacy protection systems in Large Language Models (LLMs), highlighting the limitations of current models in handling personally identifiable information (PII).
Why It Matters
As LLMs become more prevalent, ensuring user privacy is critical. This research addresses significant gaps in existing privacy protection mechanisms, providing a structured approach to evaluate and improve PII handling in AI systems, which is essential for user trust and compliance with privacy regulations.
Key Takeaways
- PII-Bench is the first comprehensive evaluation framework for query-aware privacy protection systems.
- Current LLMs perform well in basic PII detection but struggle with query relevance, especially in complex scenarios.
- The framework includes 2,842 test samples across 55 PII categories, highlighting diverse privacy challenges.
- Significant improvements are needed in intelligent PII masking to enhance user privacy.
- The research underscores the importance of robust privacy measures in AI applications.
Computer Science > Cryptography and Security arXiv:2502.18545 (cs) [Submitted on 25 Feb 2025 (v1), last revised 17 Feb 2026 (this version, v2)] Title:PII-Bench: Evaluating Query-Aware Privacy Protection Systems Authors:Hao Shen, Zhouhong Gu, Haokai Hong, Weili Han View a PDF of the paper titled PII-Bench: Evaluating Query-Aware Privacy Protection Systems, by Hao Shen and 3 other authors View PDF HTML (experimental) Abstract:The widespread adoption of Large Language Models (LLMs) has raised significant privacy concerns regarding the exposure of personally identifiable information (PII) in user prompts. To address this challenge, we propose a query-unrelated PII masking strategy and introduce PII-Bench, the first comprehensive evaluation framework for assessing privacy protection systems. PII-Bench comprises 2,842 test samples across 55 fine-grained PII categories, featuring diverse scenarios from single-subject descriptions to complex multi-party interactions. Each sample is carefully crafted with a user query, context description, and standard answer indicating query-relevant PII. Our empirical evaluation reveals that while current models perform adequately in basic PII detection, they show significant limitations in determining PII query relevance. Even state-of-the-art LLMs struggle with this task, particularly in handling complex multi-subject scenarios, indicating substantial room for improvement in achieving intelligent PII masking. Subjects: Cryptography and Security...