Skip to content

CRAfT: Constant-Round Non-Polynomial Approximation and P2P Network for Secure Transformer Inference

Authors: Zhao, Yang, Song, Hao, He, Zhang
Year: 2026
Venue: IACR ePrint 2026/1292

type: system collection: catalogue status: auto

Links: ePrint 2026/1292

Summary

Auto-extracted summary (unreviewed)

CRAfT is a secure Transformer inference system aimed at many mutually distrustful clients. It packs several users' inputs into dense RLWE ciphertexts via a trust-anchor-free P2P aggregation scheme so that linear layers need no ciphertext rotations, and it approximates sigmoid-like non-linearities (Tanh, Sigmoid, erf/GELU, and a Softmax rewritten into sigmoid form) with a branch-free trigonometric iterative fit that uses only 4 communication rounds. On BERT-base it raises multi-client throughput by up to 8.5x and cuts per-client WAN traffic by 38.4% relative to BumbleBee.

Abstract

Transformer models have recently emerged as a game-changing technology for applications in content generation. However, severe privacy concerns and the scalability bottlenecks of existing secure protocols limit their large-scale deployment. To address these challenges, we present CRAfT, an innovative system designed for highscalability and low-latency secure Transformer inference. The core contributions of CRAfT are threefold: First, we propose a novel ciphertext packing strategy tailored for multiple mutually distrustful users. This approach completely eliminates the need for ciphertext rotations in linear projections layer and effectively offloads massive communication traffic to the high-speed P2P network. Second, we introduce a novel branch-free iterative fitting strategy based on trigonometric functions, which accurately approximates sigmoidlike functions (e.g., Tanh, Sigmoid, and erf in GELU). In the interval [−3, 3], its approximation accuracy exceeds that of the plaintext Tanh-based GELU, while requiring only 4 communication rounds without relying on polynomial approximations. Finally, for complex operations like Softmax that exhibit partial sigmoid-like characteristics, we transform them into standard sigmoid-like functions, allowing us to directly apply our fitting strategy to minimize cryptographic costs while maintaining model accuracy. Comprehensive evaluations on BERT-base demonstrate that CRAfT increases multiclient inference throughput by up to 8.5× and reduces per-client WAN communication volume by 38.4% compared to the highly optimized baseline (BumbleBee).
Source: pdf

Extracted details

Field Value Source Confidence
contribution column-vector coefficient packing that eliminates ciphertext rotations in linear projections llm 0.85
contribution P2P multi-user ciphertext packaging with trust-anchor-free aggregation based on a dual common reference string llm 0.85
contribution branch-free iterative trigonometric fitting for sigmoid-like functions in 4 communication rounds without polynomial approximation llm 0.85
contribution transformation of Softmax into a standard sigmoid-like function to reuse the fitting strategy llm 0.85
schemes RLWE-based HE llm 0.85
schemes MPC llm 0.85
models BERT-base llm 0.85
tasks secure Transformer inference llm 0.85
tasks multi-client private inference llm 0.85
threat model multiple mutually distrustful clients and a server, P2P client network llm 0.85
interaction interactive (HE+MPC) llm 0.85
limitations efficiency drops when the number of clients is small because ciphertexts are sparsely packed llm 0.85
limitations dynamic statistics in Softmax and LayerNorm are replaced by static global approximations that require task-specific fine-tuning llm 0.85
limitations static approximation causes accuracy degradation, especially on fine-grained syntactic tasks llm 0.85
llm relevance high llm 0.85
key result multi-client inference throughput: up to 8.5x (BERT-base vs BumbleBee) llm 0.85
key result per-client WAN communication reduction: 38.4% (BERT-base vs BumbleBee) llm 0.85
key result communication rounds for sigmoid-like approximation: 4 (interval [-3, 3], accuracy above plaintext Tanh-based GELU) llm 0.85

System card: CRAfT

Group 5. Interactive MPC and HE+MPC hybrids
Family HE + MPC, multi-user
Arithmetic Mixed
Interaction Interactive (constant rounds)
Model scale Transformer (see paper)
Status Active SOTA
Relation to the SoK Rotation-free HE packing for many mutually distrusting users.
Note Moves communication to a peer-to-peer network to scale to many users.
Systems page CRAfT

Keyword mentions

Counts of vocabulary terms in the full text (first page of occurrence in brackets).

  • scheme: Secret sharing 16 [p.3], BFV 11 [p.3], MPC 9 [p.1], Function secret sharing 7 [p.11]
  • library: ABY / ABY3 2 [p.12], EzPC / CrypTFlow 2 [p.12], TensorFlow 1 [p.13]
  • model: Transformer 24 [p.1], BERT 6 [p.1], MLP 4 [p.9], GPT 3 [p.12], CNN 2 [p.12], LLaMA 1 [p.13], RNN / LSTM 1 [p.13]
  • dataset: GLUE tasks (QNLI/RTE/MRPC/CoLA/QQP/MNLI/STS-B) 12 [p.2], IMDB 7 [p.2], SST-2 5 [p.2], ImageNet 1 [p.12]
  • hardware: GPU 5 [p.11], NVIDIA A100 1 [p.11]
  • technique: Sigmoid 26 [p.1], GELU 24 [p.1], Softmax 18 [p.1], Polynomial approximation 15 [p.1], Two-party computation 10 [p.12], Rotations 8 [p.1], Fine-tuning 5 [p.8], LayerNorm 4 [p.2], Attention 4 [p.1], Taylor expansion 3 [p.7], SIMD packing 3 [p.2], Matrix multiplication 3 [p.3], Quantization 2 [p.3], Semi-honest model 2 [p.3]
  • assumption: RLWE 16 [p.3], LWE 2 [p.3], Lattice-based 2 [p.3]
  • topic: Homomorphic encryption 25 [p.3], Private / secure inference 11 [p.1], Privacy-preserving machine learning 9 [p.1]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/catalogue/5-136_craft-constant-round-non-polynomial-approximation-and-p2p-ne.pdf catalogue 17 998.9 KB 564a846fc053

Database key: craft-constant-round-non-polynomial-approximation-and-p2p-network-for · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set craft-constant-round-non-polynomial-approximation-and-p2p-network-for COLUMN VALUE