Skip to content

ARION: Attention-optimized transformer inference on encrypted data

Authors: L. Yang; J. Chen; W. Dai; S. Wang; W. Wu; Y. Feng
Year: 2025
Venue: IACR ePrint 2025/2271

type: system collection: catalogue, sok-references status: auto

Links: ePrint 2025/2271

Summary

Auto-extracted summary (unreviewed)

ARION is a non-interactive FHE protocol for private Transformer inference that targets the encrypted attention block. A double Baby-Step Giant-Step algorithm for the three consecutive ciphertext matrix multiplications in multi-head attention cuts rotations by 82.5% versus MOAI on BERT-Base (5.7x lower rotation cost), and a linear-nonlinear fusion evaluates softmax (shift-by-max, exponentiation, reciprocal) fused with the surrounding matrix operations. Implemented in Lattigo and evaluated on BERT-Tiny, BERT-Base and LLaMA attention kernels, it reaches plaintext-comparable classification accuracy with up to 2.5x end-to-end speedup over MOAI on BERT-Base.

Abstract

Privacy-preserving Transformer inference (PPTI) is essential for deploying large language models (LLMs) such as BERT and LLaMA in sensitive domains. In these models, the attention mechanism is both the main source of expressiveness and the dominant performance bottleneck under fully homomorphic encryption (FHE), due to large ciphertext matrix multiplications and the softmax nonlinearity. This paper presents ARION, a non-interactive FHE-based PPTI protocol that specifically optimizes the computation of encrypted attention. First, for the three consecutive ciphertext matrix multiplications in multi-head attention, we introduce the double Baby-Step Giant-Step algorithm, which significantly reduces the number of ciphertext rotations. On BERT-Base, ARION achieves an 82.5% reduction in rotations over the state-of-the-art PPTI protocol MOAI (2025), corresponding to a 5.7x reduction in rotation cost. Second, we propose a linear–nonlinear fusion technique tailored to the softmax evaluation in attention. By decomposing softmax into shift-by-maximum, exponentiation, and reciprocal sub-steps and fusing them with the surrounding encrypted matrix operations, ARION enables efficient attention evaluation while remaining compatible with diverse ciphertext packing formats. We implement ARION using Lattigo and first evaluate attention kernels on popular LLMs including BERT-Tiny, BERT-Base, and LLaMA, confirming the practicality and scalability of the proposed optimizations for encrypted attention computation. For end-to-end applications, on classification tasks for several benchmark datasets, ARION attains accuracy comparable to plaintext inference and yields up to 2.5x end-to-end speedups over MOAI for BERT-Base. 1
Source: pdf

Extracted details

Field Value Source Confidence
contribution double Baby-Step Giant-Step algorithm for the three chained ciphertext matrix multiplications in multi-head attention llm 0.85
contribution linear-nonlinear fusion technique for softmax compatible with diverse packing formats llm 0.85
contribution open-source end-to-end non-interactive FHE Transformer inference implementation in Lattigo llm 0.85
contribution attention-kernel evaluation on BERT-Tiny, BERT-Base and LLaMA and end-to-end classification benchmarks llm 0.85
schemes CKKS llm 0.85
libraries Lattigo llm 0.85
models BERT-Tiny llm 0.85
models BERT-base llm 0.85
models LLaMA llm 0.85
tasks private transformer inference llm 0.85
tasks text classification llm 0.85
threat model client-server, single untrusted server (non-interactive) llm 0.85
interaction non-interactive FHE llm 0.85
code url https://github.com/hangenba/Arion.learning (URL found in the full text) regex 0.50
code url https://github.com/microsoft/SEAL (URL found in the full text) regex 0.50
code url https://github.com/Desilo/liberate-fhe (URL found in the full text) regex 0.50
code url https://github.com/hangenba/Arion llm 0.85
llm relevance high llm 0.85
key result ciphertext rotation reduction: 82.5% fewer rotations than MOAI (BERT-Base attention) llm 0.85
key result rotation cost reduction: 5.7x (BERT-Base vs MOAI) llm 0.85
key result end-to-end speedup: up to 2.5x over MOAI (BERT-Base classification tasks) llm 0.85
key result accuracy: comparable to plaintext inference (classification on several benchmark datasets) llm 0.85

System card: ARION

Group 0. CKKS LLM frameworks surveyed in the SoK
Family CKKS · Model-modifying (N-MOD)
Arithmetic Approximate real (CKKS)
Interaction non-interactive FHE
Model scale BERT-Tiny, BERT-Base
Library Lattigo
Hardware 64-core CPU
Packing L-COL
Latency 9.3 s (Tiny) / 225 s (Base) amortized per sequence, batch 256
Open source yes
Status Active SOTA
Relation to the SoK CPU-only Lattigo implementation with column/diagonal packing.
Note Strong amortized throughput on CPU; requires architectural modifications.
Systems page ARION

SoK reference [13]

L. Yang, J. Chen, W. Dai, S. Wang, W. Wu, and Y. Feng, “ARION: Attention-optimized transformer inference on encrypted data,” Cryptology ePrint Archive, Report 2025/2271, 2025. [Online]. Available: https://eprint.iacr.org/2025/2271

Listed on the reference page of the site.

Keyword mentions

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

  • scheme: CKKS 28 [p.3], MPC 15 [p.1], Secret sharing 2 [p.2]
  • library: Lattigo 26 [p.1], Microsoft SEAL 18 [p.3], HELayers 4 [p.3], OpenFHE 3 [p.3], HElib 3 [p.13], Phantom 3 [p.3], Liberate.FHE 3 [p.3]
  • model: BERT 84 [p.1], Transformer 27 [p.1], LLaMA 12 [p.1], ViT 3 [p.13], GPT 2 [p.1], CNN 1 [p.12]
  • dataset: GLUE tasks (QNLI/RTE/MRPC/CoLA/QQP/MNLI/STS-B) 12 [p.7], SST-2 5 [p.10]
  • hardware: GPU 13 [p.1], Intel Xeon 1 [p.10], CUDA 1 [p.12]
  • technique: Rotations 50 [p.1], Softmax 37 [p.1], Matrix multiplication 26 [p.1], LayerNorm 22 [p.2], Bootstrapping 21 [p.2], GELU 16 [p.2], Non-interactive 12 [p.1], Attention 9 [p.1], Chebyshev 8 [p.8], Semi-honest model 8 [p.5], SIMD packing 6 [p.1], Fine-tuning 4 [p.10], Key switching 3 [p.14], Polynomial approximation 3 [p.8], Malicious model 3 [p.5], Goldschmidt 2 [p.8], Newton iteration 2 [p.9], Convolution 2 [p.12]
  • assumption: Lattice-based 5 [p.10], RLWE 4 [p.4], IND-CPA 3 [p.9], LWE 2 [p.13]
  • task: Genomics 2 [p.2]
  • topic: Homomorphic encryption 59 [p.1], Private / secure inference 22 [p.1], Large language models 11 [p.1], Benchmark 9 [p.1], Open source 9 [p.1], Privacy-preserving machine learning 6 [p.1]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/catalogue/0-006_arion-attention-optimized-transformer-inference-on-encrypted.pdf catalogue 17 609.9 KB 337351faeef6
references/sok-references/013_arion-attention-optimized-transformer-inference-on-encrypted.pdf sok-references 17 609.9 KB 337351faeef6 byte-identical to references/catalogue/0-006_arion-attention-optimized-transformer-inference-on-encrypted.pdf

Database key: arion-attention-optimized-transformer-inference-on-encrypted-data · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set arion-attention-optimized-transformer-inference-on-encrypted-data COLUMN VALUE