Dice Question Streamline Icon: https://streamlinehq.com

Scalable verifiable audit mechanisms with low overhead

Develop verifiable audit architectures for AI inference pipelines that do not require performing all evaluation and inference inside trusted execution environments and that minimize computational overhead to enable broad adoption.

Information Square Streamline Icon: https://streamlinehq.com

Background

A proposed two‑stage verification approach relies on TEEs to attest to evaluated pipelines, but significant overhead limits scalability.

Reducing overhead and enabling broader, more scalable verification would allow wider use of verifiable audits across diverse systems.

References

Another open problem is that this pipeline requires that all evaluation and inference is done in enclaves and with significant computational overhead, effectively limiting verifiable audits for a few critical systems, and necessitating more scalable structures for verifying audits.

— Open Problems in Technical AI Governance (2407.14981 - Reuel et al., 20 Jul 2024) in Section 5.4.1 “Verifiable Audits”