BlindStep MPC kernel for privacy-preserving collaborative reasoning
Original title: BlindStep: A Secure One-Step Kernel for Privacy-Preserving Collaborative Reasoning
BlindStep is a fixed-shape MPC-online kernel that provides simulation-based guarantees for bounded-slot aggregation between semi-honest parties.
BlindStep is presented as a secure one-step kernel for post-alignment steps in two-party collaborative workflows. It realizes six semi2k variants in MP-SPDZ across sharing styles and top-K strategies. Performance tests show B-Queue achieving major speedups and communication reductions over H-variants under LAN, while H-Sort performs best in WAN scenarios due to round complexity.
- Six semi2k realizations implemented in MP-SPDZ
- B-Queue 6-7 times faster than H-Queue at N=256 under LAN
- B-Queue delivers 21-30x speedups over H-Sort at N=2048-8192
- H-Sort fastest under N=256 WAN profile
From the original article
1 min read at arxiv.orgA recurring step in two-party collaborative workflows combines a private candidate frontier with a partner's compatible evidence, scores the resulting candidates, and retains only a small frontier. Direct execution can expose intermediate candidates, local scores, or unauthorized evidence provenance.
By Xiaoxian Liu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.