Caliber defense mitigates AI model extraction via output perturbation

Researchers have developed Caliber, a novel defense mechanism against model extraction attacks on score-returning APIs. Caliber works by adding Gaussian noise to internal logits, which degrades the supervision signal used to train surrogate models. The system provides provable per-input query costs for recovering clean logits and demonstrates a mean absolute relative error of 0.6-1.4% across numerous model-dataset combinations. AI

IMPACT Introduces a new method to protect AI models from extraction, potentially increasing security for API providers.

RANK_REASON The cluster contains a research paper detailing a new defense mechanism against model extraction. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• Caliber
• CatalyzeX
• DagsHub
• Gaussian function
• Gotit.pub
• Hugging Face
• IArxiv
• ScienceCast

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →