Reproduce “Unmasking Database Vulnerabilities: Zero-Knowledge Schema Inference Attacks in Text-to-SQL Systems” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Đorđe Klisura, Anthony Rios
A zero-knowledge framework, developed by The University of Texas at San Antonio, unmasks sensitive database schema information from text-to-SQL systems by iteratively probing for table and column details. This method consistently achieves high F1 scores for schema reconstruction, revealing a critical vulnerability and the generalizability of "prompt stealing" attacks against LLMs.
Key insights
- Text-to-SQL systems are highly vulnerable to zero-knowledge schema inference attacks, enabling attackers to reconstruct detailed database structures without prior information.
- Simple prompt-based protection mechanisms are largely ineffective in preventing substantial schema leakage from generative LLMs in text-to-SQL contexts.
- The identified vulnerability extends beyond text-to-SQL, demonstrating a broader "prompt stealing" risk where proprietary prompt information can be extracted from other LLM applications.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.