A Graph-Based Analysis of Comparability, Scale, and Attribution in Private Transformer Inference
A cs.CR study graphs 62 private Transformer inference systems and finds almost none are directly comparable.
This cs.CR paper systematizes private Transformer inference by building an empirical graph of 62 systems, mapping attributes to nodes and evaluation compatibility to edges. Node-level attribution finds that cryptographic progress drives most reported gains, but always alongside changes in model fidelity, hardware, or cost accounting. Edge auditing finds only 1 of 1,891 candidate pairs directly comparable under shared assumptions, and the scaling frontier of private inference has plateaued at 13 billion parameters.
- Authors graph 62 private Transformer inference systems and their evaluation compatibility.
- Only 1 of 1,891 candidate system pairs is directly comparable under shared assumptions.
- Cryptographic speedups always co-occur with fidelity, hardware, or cost-accounting changes.
- The private-inference scaling frontier has plateaued at about 13 billion parameters.
From the original article
1 min read at arxiv.orgRising privacy concerns around large language models have accelerated research in private Transformer inference (PTI). Synthesizing the empirical claims in this line of research, however, is hindered by multi-dimensional cryptographic progress and inconsistent evaluation contracts.
By Souvik Das · 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.