Skip to content

Back to the catalog

A Graph-Based Analysis of Comparability, Scale, and Attribution in Private Transformer Inference
eprint.iacr.orgpaper

A Graph-Based Analysis of Comparability, Scale, and Attribution in Private Transformer Inference

Systematization of knowledge on private Transformer inference (PTI) that builds an empirical graph of 62 systems, with system attributes as node features and evaluation compatibility as edges. Node-level attribution finds that cryptographic progress drives most speedups but always co-occurs with confounders such as model fidelity, hardware and cost accounting. Edge-level auditing finds only 1 of 1891 candidate system pairs directly comparable. Graph-level analysis shows the scale frontier plateaued at 13B parameters and projects the cost of deploying current PTI techniques on production-scale Transformers.

TL;DR SoK graph analysis of 62 MPC/HE private Transformer inference systems: few results are comparable (1 of 1891 pairs), and scale has plateaued at 13B parameters.
Open original ↗
Document

What it claims

This is a systematization-of-knowledge preprint on private Transformer inference (PTI). It covers systems built on secure multi-party computation (MPC) and homomorphic encryption (HE), the two primitives listed in its keywords. The authors encode 62 published PTI systems as a graph. Each system's attributes become node features, and an edge records whether two systems' evaluation setups are compatible.

The headline audit number is easy to misread. 1891 is exactly 62×61/2, so it is every possible pair in the corpus, not a pre-filtered candidate set. Only one pair among all of them can be compared directly under shared assumptions. In practice, a speedup that one PTI paper reports over another almost never comes from a like-for-like comparison.

The attribution finding matters for anyone choosing a protocol. Cryptographic improvements account for most reported gains, but in every case they arrive bundled with at least one confounder:

  • changes to model fidelity, such as approximated non-linearities or altered models,
  • different hardware,
  • different cost accounting, meaning what is counted as latency or communication.

A headline speedup therefore cannot be credited to the protocol alone.

The paper frames the field's difficulty as two problems: progress along many cryptographic dimensions at once, and inconsistent "evaluation contracts" between papers. It offers the graph framework as a way for others to check PTI claims transparently.

How to use it

  • Read it before trusting cross-paper speedup tables in MPC/HE inference work, or before picking a PTI system based on another paper's reported baselines.
  • Licence: CC BY.
  • Status: preprint on the IACR ePrint archive (Applications category). Received 2026-10-13, approved 2026-10-14. Short URL: https://ia.cr/2026/2464.

Limits

  • Only the abstract and metadata were available, so this document is limited to what the abstract states.
  • The abstract does not name:
  • the single comparable pair,
  • which systems reach the 13B-parameter frontier,
  • the projected deployment costs for production-scale models, or the model sizes used in that projection.
  • It also does not say whether the graph dataset or analysis code is released. Check the full PDF for that before planning to build on the framework.
  • The corpus is 62 systems as of this preprint. Systems published after its cutoff are not covered.

Organisations named here

  • Texas A&M University
  • Ariful Azad
  • Nitesh Saxena
  • Souvik Das

Technologies named here

  • LLM

Techniques named here

  • Private Transformer Inference
  • Empirical graph analysis of PTI systems

Also named here

  • Transformer

Related in research