Euston: Efficient and User-Friendly Secure Transformer Inference with Non-Interactivity¶
Authors: X. Gao; S. Fu; L. Liu; Z. Liu; Y. Luo; Y. Wang
Year: 2026
Venue: IACR ePrint 2026/046
type: system collection: catalogue, sok-references status: auto
Links: ePrint 2026/46 · link
Summary¶
Auto-extracted summary (unreviewed)
Euston is a non-interactive RNS-CKKS secure transformer inference framework (built on the NEXUS codebase for CPU via SEAL and GPU via Phantom) that uses SVD-based weight handling with batched homomorphic matrix multiplications to shrink ciphertexts, rotations and user-side preprocessing, and column/diagonal-packed layouts plus depth-regulation strategies for rotation-free, shallower nonlinear evaluations. Against NEXUS it reports up to 3100x lower user preprocessing cost, 8.8x higher system-wide inference performance, 90x faster HMM (CPU) and 165.7x faster HNE, with accuracy within a fraction of a percent of plaintext for BERT-base and LLaMA-3-8B on RTE, SST-2 and QNLI.
Abstract¶
Secure TransFormer Inference (STFI) frameworks have been proposed to address privacy concerns over user inputs and model parameters in Transformer-based LLMs. While most existing solutions rely on interactive protocols that incur substantial user-server communication overhead, non-interactive STFI variants have recently emerged to eliminate such dependencies. Nevertheless, state-of-the-art noninteractive STFI frameworks still suffer from critical limitations. (i) Large ciphertext sizes and multiple rotations alongside heavy user-side overhead in Homomorphic Matrix Multiplication (HMM). (ii) High approximation costs and depth consumptions in Homomorphic Nonlinear Evaluations (HNE). To address these limitations, we present Euston, an efficient and user-friendly STFI with non-interactivity. By combining RNS-CKKS fully homomorphic encryption with optimized methods, Euston achieves unprecedented efficiency in offlineonline inference paradigm. The key innovations are twofold. (i) For linear operations, we adopt Singular Value Decomposition (SVD) with our novel batched HMMs to minimize ciphertext size and reduce rotation counts, simultaneously lowering user-side computational, communication and storage overhead. (ii) For nonlinear operations, we employ column(diagonal)packed ciphertext matrix formats to eliminate costly rotations and depth regulation strategies to reduce depth consumption in non-interactive HNEs, which not only avoids user-server communications but also accelerates inference performance. In comparision with the state-of-the-art approach (NEXUS, NDSS 2025), Euston achieves up to 3100× lower preprocessing costs for the user and 8.8× higher system-wide inference performance, specifically delivering a 90× speedup for HMM and a 165.7× speedup for HNE. Our results demonstrate that Euston establishes new efficiency frontiers for user-friendly STFI deployment across cloud and edge environments.
Source: pdf
Extracted details¶
| Field | Value | Source | Confidence |
|---|---|---|---|
| contribution | SVD-based secure weight transmission with batched homomorphic matrix multiplication reducing ciphertext size and rotations | llm | 0.85 |
| contribution | Column/diagonal-packed ciphertext formats eliminating rotations in nonlinear evaluation | llm | 0.85 |
| contribution | Depth regulation strategies reducing multiplicative depth of non-interactive nonlinear evaluations | llm | 0.85 |
| contribution | CPU (SEAL) and GPU (Phantom) implementation and comparison with NEXUS on BERT-base and LLaMA-3-8B | llm | 0.85 |
| schemes | CKKS | llm | 0.85 |
| libraries | SEAL | llm | 0.85 |
| libraries | Phantom | llm | 0.85 |
| models | BERT-base | llm | 0.85 |
| models | LLaMA-3-8B | llm | 0.85 |
| datasets | RTE | llm | 0.85 |
| datasets | SST-2 | llm | 0.85 |
| datasets | QNLI | llm | 0.85 |
| tasks | private transformer inference | llm | 0.85 |
| tasks | GLUE text classification | llm | 0.85 |
| hardware | AMD EPYC 7542 CPU @ 2.90 GHz, 32 threads, 128 GB RAM | llm | 0.85 |
| hardware | NVIDIA RTX 6000 Ada 48GB GPU | llm | 0.85 |
| security level | 128-bit | llm | 0.85 |
| threat model | client-server, offline-online, non-interactive; protects user inputs and model parameters | llm | 0.85 |
| interaction | non-interactive FHE | llm | 0.85 |
| code url | https://github.com/FLL-Lab/Euston (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/encryptorion-lab/phantom-fhe (URL found in the full text) | regex | 0.50 |
| limitations | NEXUS releases only partial code, so some comparisons rely on published results | llm | 0.85 |
| limitations | Only a LAN network setting is simulated | llm | 0.85 |
| llm relevance | high | llm | 0.85 |
| key result | user-side preprocessing cost reduction: up to 3100x (vs NEXUS) | llm | 0.85 |
| key result | system-wide inference speed-up: 8.8x (vs NEXUS) | llm | 0.85 |
| key result | HMM speed-up: 90x CPU, 9.2x GPU, 2.8x less communication (128x768 by 768x768 input-weight matmul vs NEXUS, small batch) | llm | 0.85 |
| key result | HNE speed-up: 165.7x (nonlinear evaluations vs NEXUS) | llm | 0.85 |
| key result | accuracy: 93.12% (plaintext 93.23%, NEXUS 92.98%) (BERT-base on SST-2) | llm | 0.85 |
| key result | accuracy: 81.56% (plaintext 82.75%, NEXUS 81.24%) (LLaMA-3-8B on RTE) | llm | 0.85 |
System card: EUSTON¶
| 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-Base, GPT-2-1.5B, Llama-3-8B |
| Library | SEAL + Phantom |
| Hardware | RTX 6000 GPU |
| Packing | L-ALG |
| Latency | 46 s (BERT-Base) / 136 s (GPT-2-1.5B) / 484 s (Llama-3-8B) amortized, batch 32 |
| Open source | yes |
| Status | Active SOTA |
| Relation to the SoK | SVD/precomputed transforms to reduce depth; NEXUS arg max. |
| Note | Scales to Llama-3-8B classification with modest accuracy loss. |
| Systems page | EUSTON |
SoK reference [19]¶
X. Gao, S. Fu, L. Liu, Z. Liu, Y. Luo, and Y. Wang, “Euston: Efficient and user-friendly secure transformer inference with noninteractivity,” Cryptology ePrint Archive, 2026.
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: MPC 9 [p.1], CKKS 8 [p.1], Secret sharing 2 [p.15], Function secret sharing 2 [p.15]
- library: Phantom 2 [p.11], EzPC / CrypTFlow 2 [p.14], Microsoft SEAL 1 [p.11]
- model: Transformer 16 [p.1], BERT 15 [p.2], GPT 9 [p.14], RNN / LSTM 2 [p.14], CNN 1 [p.16]
- dataset: GLUE tasks (QNLI/RTE/MRPC/CoLA/QQP/MNLI/STS-B) 6 [p.12], SST-2 3 [p.12], GLUE 2 [p.12], ImageNet 1 [p.14]
- hardware: GPU 18 [p.2], NVIDIA RTX 1 [p.12], AMD EPYC 1 [p.12], CUDA 1 [p.12]
- technique: Rotations 30 [p.1], GELU 25 [p.2], Softmax 22 [p.2], LayerNorm 21 [p.2], Matrix multiplication 21 [p.1], Bootstrapping 16 [p.2], Non-interactive 14 [p.1], RNS 8 [p.1], Attention 6 [p.1], SIMD packing 6 [p.2], Two-party computation 5 [p.14], Semi-honest model 4 [p.4], Polynomial approximation 3 [p.4], Goldschmidt 2 [p.9], Sigmoid 2 [p.10]
- assumption: IND-CPA 3 [p.6]
- task: Finance 2 [p.12]
- topic: Homomorphic encryption 33 [p.1], Private / secure inference 9 [p.4], Benchmark 7 [p.2], Large language models 5 [p.1], Open source 2 [p.11]
Files¶
| Path (under papers/) | Collection | Pages | Size | SHA-256 | Note |
|---|---|---|---|---|---|
references/catalogue/0-012_euston-efficient-and-user-friendly-secure-transformer-infere.pdf |
catalogue | 20 | 3.9 MB | 2e5e8bcdd251 |
|
references/sok-references/019_euston-efficient-and-user-friendly-secure-transformer-infere.pdf |
sok-references | 20 | 3.9 MB | 2e5e8bcdd251 |
byte-identical to references/catalogue/0-012_euston-efficient-and-user-friendly-secure-transformer-infere.pdf |
Database key: euston-efficient-and-user-friendly-secure-transformer-inference-with · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set euston-efficient-and-user-friendly-secure-transformer-inference-with COLUMN VALUE