Reproduce “OML: A Primitive for Reconciling Open Access with Owner Control in AI Model Distribution” 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
Zerui Cheng, Edoardo Contente, Ben Finch, Oleg Golev, Jonathan Hayase, Andrew Miller, Niusha Moshrefi, Anshul Nasery and 4 more
This research introduces OML, a primitive that reconciles open access to AI models with owner control, monetization, and policy enforcement. The framework enables local model execution while allowing creators to capture economic value and technically enforce ethical guidelines, demonstrating its feasibility with practical fingerprinting techniques and theoretical underpinnings.
Key insights
- OML formally defines a locally executable model that produces high-utility outputs only when presented with valid, cryptographically-bound permissions, even under white-box access conditions.
- Theoretical results establish an information-theoretic impossibility for perfect security against unbounded adversaries, but prove feasibility if indistinguishability obfuscation exists, and characterize a query-security trade-off.
- The design space for OML constructions encompasses various approaches, including software obfuscation, hardware-backed Trusted Execution Environments (TEEs), cryptographic encryption (FHE/FE), and hybrid compositions, each offering distinct security-performance trade-offs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.