Secure transformer inference made non-interactive¶
Authors: J. Zhang; X. Yang; L. He; K. Chen; W. jie Lu; Y. Wang; X. Hou; J. Liu; K. Ren; X. Yang
Year: 2025
Venue: NDSS 2025
type: system collection: catalogue, sok-references status: auto
Links: DOI · ePrint 2024/136
Summary¶
Auto-extracted summary (unreviewed)
NEXUS is presented as the first non-interactive protocol for secure transformer inference: the client submits one encrypted input and receives one encrypted result, with no further rounds. New primitives (SIMD ciphertext compression/decompression, SIMD slot folding, secure Argmax) cut bandwidth by 372.5x versus BOLT and 53.6x versus Bumblebee at comparable runtime, and the non-interactive design allows a GPU version with a 42.3x runtime speedup, running a BERT-based model in 37.3 s with 164 MB of communication.
Abstract¶
Secure transformer inference has emerged as a prominent research topic following the proliferation of ChatGPT. Existing solutions are typically interactive, involving substantial communication load and numerous interaction rounds between the client and the server. In this paper, we propose NEXUS, the first non-interactive protocol for secure transformer inference. The protocol requires the client to engage in just one round of communication with the server during the whole inference process: submitting an encrypted input and receiving an encrypted result. NEXUS introduces several novel primitives, including SIMD ciphertext compression/decompression, SIMD slot folding, and secure Argmax, which enable it to significantly surpass the state-of-the-art in communication while maintaining comparable runtime. Specifically, it reduces bandwidth consumption by 372.5× compared to BOLT (Oakland ’24) and 53.6× compared to Bumblebee (NDSS ’25). Furthermore, its non-interactive property allows for optimal hardware acceleration, with the GPU version achieving a 42.3× speedup in runtime. This enables NEXUS to run inference on a BERT-based model in just 37.3 seconds, consuming only 164 MB of bandwidth.
Source: pdf
Extracted details¶
| Field | Value | Source | Confidence |
|---|---|---|---|
| contribution | First non-interactive (single-round) secure transformer inference protocol | llm | 0.75 |
| contribution | SIMD ciphertext compression/decompression and SIMD slot folding primitives | llm | 0.75 |
| contribution | Secure Argmax primitive | llm | 0.75 |
| contribution | GPU-accelerated implementation exploiting the non-interactive design; open-sourced code | llm | 0.75 |
| schemes | CKKS | llm | 0.75 |
| libraries | SEAL | llm | 0.75 |
| libraries | FHE-MP-CNN | llm | 0.75 |
| libraries | Phantom | llm | 0.75 |
| models | BERT-based model | llm | 0.75 |
| tasks | private Transformer inference | llm | 0.75 |
| hardware | GPU (model not specified in excerpt) | llm | 0.75 |
| security level | 128-bit | llm | 0.75 |
| threat model | two-party client-server; server learns nothing about the input, client learns nothing about the model beyond the result | llm | 0.75 |
| interaction | non-interactive FHE | llm | 0.75 |
| code url | https://github.com/zju-abclab/NEXUS.Figure (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/snu-ccl/FHE-MP-CNN (URL found in the full text) | regex | 0.50 |
| code url | https://github.com/zju-abclab/NEXUS | llm | 0.75 |
| llm relevance | high | llm | 0.75 |
| key result | bandwidth reduction: 372.5x (vs BOLT (Oakland '24)) | llm | 0.75 |
| key result | bandwidth reduction: 53.6x (vs Bumblebee (NDSS '25)) | llm | 0.75 |
| key result | GPU runtime speedup: 42.3x (GPU version vs CPU version) | llm | 0.75 |
| key result | end-to-end inference latency: 37.3 s (BERT-based model, GPU) | llm | 0.75 |
| key result | communication: 164 MB (BERT-based model inference) | llm | 0.75 |
| key result | estimated cost per reply token: $0.05 (17.2x reduction) (AWS pricing, vs BOLT $5.44) | llm | 0.75 |
System card: NEXUS¶
| 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, Llama-3-8B |
| Library | SEAL + Phantom |
| Hardware | 4x A100 GPUs |
| Packing | L-CYC |
| Latency | 37 s BERT-Base amortized (batch 32); 52 s/token Llama-3-8B |
| Open source | yes |
| Status | Active SOTA |
| Relation to the SoK | Introduced O(log V) SIMD-folding QuickMax arg max adopted by CEP, MOAI, EUSTON. |
| Note | First end-to-end non-interactive CKKS transformer; widely used baseline. |
| Code | https://github.com/zju-abclab/NEXUS |
| Systems page | NEXUS |
SoK reference [15]¶
J. Zhang, X. Yang, L. He, K. Chen, W. jie Lu, Y. Wang, X. Hou, J. Liu, K. Ren, and X. Yang, “Secure transformer inference made non-interactive.” in NDSS. The Internet Society, 2025. [Online]. Available: http://dblp.uni-trier.de/db/conf/ndss/ndss2025. html#ZhangYH0LWH00025
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 18 [p.2], MPC 8 [p.12], Secret sharing 3 [p.13], RSA 2 [p.14]
- library: Microsoft SEAL 3 [p.9], SecretFlow-SPU 3 [p.10]
- model: BERT 21 [p.1], CNN 10 [p.1], LLaMA 8 [p.2], GPT 5 [p.1], Transformer 4 [p.1], RNN / LSTM 1 [p.15]
- dataset: GLUE tasks (QNLI/RTE/MRPC/CoLA/QQP/MNLI/STS-B) 7 [p.12], SST-2 3 [p.12], GLUE 2 [p.12], ImageNet 1 [p.2]
- hardware: GPU 15 [p.1], NVIDIA A100 2 [p.10], FPGA 1 [p.1], Intel Xeon 1 [p.10], Multi-GPU 1 [p.14]
- technique: Bootstrapping 29 [p.3], GELU 25 [p.2], SIMD packing 25 [p.1], Matrix multiplication 23 [p.1], Non-interactive 20 [p.1], RNS 18 [p.2], LayerNorm 16 [p.2], Softmax 14 [p.2], Rotations 14 [p.2], Two-party computation 6 [p.1], Membership inference 6 [p.2], ReLU 4 [p.13], Key switching 3 [p.5], Attention 3 [p.2], Semi-honest model 3 [p.2], Compiler 3 [p.14], Polynomial approximation 2 [p.4], Goldschmidt 2 [p.8], Convolution 2 [p.1]
- assumption: RLWE 3 [p.16]
- task: Finance 7 [p.1]
- topic: Homomorphic encryption 40 [p.2], Private / secure inference 17 [p.1], Benchmark 10 [p.10], Open source 5 [p.2]
Files¶
| Path (under papers/) | Collection | Pages | Size | SHA-256 | Note |
|---|---|---|---|---|---|
references/catalogue/0-015_secure-transformer-inference-made-non-interactive.pdf |
catalogue | 17 | 1.7 MB | d4f125addb2b |
|
references/sok-references/015_secure-transformer-inference-made-non-interactive.pdf |
sok-references | 17 | 1.7 MB | d4f125addb2b |
byte-identical to references/catalogue/0-015_secure-transformer-inference-made-non-interactive.pdf |
Database key: secure-transformer-inference-made-non-interactive · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set secure-transformer-inference-made-non-interactive COLUMN VALUE