analiza aiqbench
This research introduces "Knowledge Trap," a novel defense mechanism that mitigates large language model extraction attacks by redirecting adversaries toward low-value knowledge nodes within a honeypot graph. By consuming attacker query budgets on negligible utility data, the method significantly reduces surrogate model agreement while preserving accuracy for legitimate users in sensitive domains like medicine and finance. This approach represents a practical shift from reactive blocking to proactive resource depletion in the security of commercial AI APIs.
Let Them Steal: Trapping Large Language Model Extraction Attacks with Knowledge Honeypot
arXiv:2606.15810v2 Announce Type: replace-cross Abstract: Large language models deployed as commercial APIs are vulnerable to model extraction attacks, while existing defenses either act too late or degrade utility for legitimate users. We propose \textbf{Knowledge Trap}, a defense that redirects extraction attacks toward low-transferability knowledge through a \emph{Honeypot Knowledge Graph} (HKG) and breadcrumb-guided exploration. Instead of blocking queries or perturbing outputs, Knowledge Trap consumes the attacker's limited query budget on knowledge with negligible downstream utility while preserving benign-user performance. Experiments in medical and financial domains show that Knowledge Trap reduces surrogate Agreement by 6.2\% on average without degrading legitimate-user accuracy, outperforming existing defenses that impose measurable user impact. These results suggest that defending knowledge-space traversal is a practical direction for mitigating LLM extraction attacks.
Abstract:Large language models deployed as commercial APIs are vulnerable to model extraction attacks, while existing defenses either act too late or degrade utility for legitimate users. We propose \textbf{Knowledge Trap}, a defense that redirects extraction attacks toward low-transferability knowledge through a \emph{Honeypot Knowledge Graph} (HKG) and breadcrumb-guided exploration. Instead of blocking queries or perturbing outputs, Knowledge Trap consumes the attacker's limited query budget on knowledge with negligible downstream utility while preserving benign-user performance. Experiments in medical and financial domains show that Knowledge Trap reduces surrogate Agreement by 6.2\% on average without degrading legitimate-user accuracy, outperforming existing defenses that impose measurable user impact. These results suggest that defending knowledge-space traversal is a practical direction for mitigating LLM extraction attacks.
Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
arXiv:2606.15810 [cs.CR]
(or arXiv:2606.15810v2 [cs.CR] for this version)
https://doi.org/10.48550/arXiv.2606.15810 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 214 cuvinte
Subiecte conexe
Let Them Steal: Trapping Large Language Model Extraction Attacks with Knowledge Honeypot
arXiv · 16 iun. 2026
Links #5: 2026/07
LessWrong · 31 iul. 2026
Distillation Defenses Easily Break After Reinforcement Learning
arXiv · 30 sept. 2026
Reward Stealing Attack on Large Language Models
arXiv · 6 oct. 2026