Knowledge Honeypots and Query-Budget Defense against Language-Model Extraction

Abstract

Llm Extraction Defense has become a test of how well researchers can connect performance with evidence quality, resource limits, and transfer across settings. This methodological synthesis evaluates redirecting adversarial exploration while preserving utility for benign users. Its analysis connects 1 focal paper with 13 independently retrieved publications confirmed at bibliographic registration or publisher level. The analysis is organized around attack modeling, honeypot knowledge, query economics, utility preservation, and adaptive attackers. Rather than treating reported outcomes as directly interchangeable, the review compares research framing, method assumptions, and test envelope. Across the literature, the comparison suggests that advances in LLM extraction defense become credible when representation, objective, and evaluation protocol are evaluated together and when uncertainty about distribution shift is reported explicitly. The framework consequently connects method selection to operational consequence while identifying external-validity hazards, and proposes a research agenda centered on well-specified controls, robustness tests, and reproducible workflows.