Skip to content

BumbleBee: Secure Two-party Inference Framework for Large Transformers

Authors: Lu, Huang, Gu, Li et al.
Year: 2025
Venue: NDSS 2025

type: system collection: catalogue status: auto

Links: DOI · ePrint 2023/1678

Summary

Auto-extracted summary (unreviewed)

BumbleBee is a two-party private inference framework for large transformers where the client holds the input and the server the model, combining RLWE-based HE (SEAL) with secret-sharing/OT protocols. It introduces matrix-multiplication protocols that cut communication by 80-90% and tailored protocols for transformer activations (cutting communication by 80-95%), is benchmarked on five transformer models including LLaMA-7B (one token in about 14 minutes on CPUs), and is over 10x faster than Iron and 3x faster than BOLT with one-tenth of BOLT's communication.

Abstract

Large transformer-based models have realized stateof-the-art performance on lots of real-world tasks such as natural language processing and computer vision. However, with the increasing sensitivity of the data and tasks they handle, privacy has become a major concern during model deployment. In this work, we focus on private inference in two-party settings, where one party holds private inputs and the other holds the model. We introduce BumbleBee, a fast and communication-friendly twoparty private transformer inference system. Our contributions are three-fold: First, we propose optimized protocols for matrix multiplication, which significantly reduce communication costs by 80% – 90% compared to previous techniques. Secondly, we develop a methodology for constructing efficient protocols tailored to the non-linear activation functions employed in transformer models. The proposed activation protocols have realized a significant enhancement in processing speed, alongside a remarkable reduction in communication costs by 80% – 95% compared with two prior methods. Lastly, we have performed extensive benchmarks on five transformer models. BumbleBee demonstrates its capability by evaluating the LLaMA-7B model, generating one token in approximately 14 minutes using CPUs. Our results further reveal that BumbleBee outperforms Iron (NeurIPS22) by over an order of magnitude and is three times faster than BOLT (Oakland24) with one-tenth communication.
Source: pdf

Extracted details

Field Value Source Confidence
contribution optimized 2PC matrix multiplication protocols reducing communication by 80-90% llm 0.80
contribution methodology for efficient protocols for transformer non-linear activations with 80-95% less communication llm 0.80
contribution benchmarks on five transformer models including LLaMA-7B private token generation llm 0.80
contribution over 10x speed-up over Iron and 3x over BOLT with one-tenth communication llm 0.80
contribution open-source implementation (OpenBumbleBee) llm 0.80
schemes BFV llm 0.80
schemes secret sharing llm 0.80
schemes oblivious transfer llm 0.80
libraries SEAL llm 0.80
libraries EMP-toolkit llm 0.80
libraries YACL llm 0.80
models LLaMA-7B llm 0.80
tasks private transformer inference llm 0.80
tasks private LLM token generation llm 0.80
hardware CPU llm 0.80
threat model two-party: client holds private input, server holds model (same security level as KRDY protocols) llm 0.80
interaction interactive (HE+MPC) llm 0.80
code url https://github.com/AntCPLab/OpenBumbleBee (URL found in the full text) regex 0.50
code url https://github.com/AntCPLab/OpenBumbleBee.A (URL found in the full text) regex 0.50
code url https://github.com/Clive2312/EzPC/tree/bert/SCI (URL found in the full text) regex 0.50
code url https://github.com/AntCPLab/OpenBumbleBee llm 0.80
llm relevance high llm 0.80
key result matrix multiplication communication reduction: 80-90% (vs previous 2PC techniques) llm 0.80
key result activation protocol communication reduction: 80-95% (vs two prior methods) llm 0.80
key result latency per generated token: about 14 minutes (LLaMA-7B private inference on CPUs) llm 0.80
key result speed-up over Iron: more than 10x (private transformer inference) llm 0.80
key result speed-up and communication vs BOLT: 3x faster with one-tenth communication (private transformer inference) llm 0.80

System card: BumbleBee

Group 5. Interactive MPC and HE+MPC hybrids
Family RLWE HE + OT 2PC
Arithmetic Fixed point
Interaction Interactive 2PC
Model scale LLaMA-7B
Status Active SOTA
Relation to the SoK Hybrid 2PC with efficient HE matmul and accurate activations.
Note Runs LLaMA-7B in about 14 minutes per token. Model-preserving.
Code https://github.com/AntCPLab/OpenBumbleBee
Systems page BumbleBee

Keyword mentions

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

  • scheme: MPC 20 [p.1], Secret sharing 10 [p.3], Oblivious transfer 3 [p.3]
  • library: SecretFlow-SPU 8 [p.3], Microsoft SEAL 5 [p.11], EzPC / CrypTFlow 5 [p.10], Concrete (Zama) 2 [p.11], ABY / ABY3 2 [p.14], HElib 1 [p.14], PyTorch 1 [p.15], HuggingFace 1 [p.3]
  • model: Transformer 22 [p.1], BERT 20 [p.1], ViT 18 [p.1], GPT 16 [p.1], LLaMA 10 [p.1], CNN 4 [p.1], RNN / LSTM 1 [p.15]
  • dataset: GLUE tasks (QNLI/RTE/MRPC/CoLA/QQP/MNLI/STS-B) 7 [p.11], ImageNet 6 [p.11], GLUE 4 [p.11]
  • hardware: GPU 3 [p.2], CUDA 1 [p.15]
  • technique: GELU 43 [p.2], SIMD packing 31 [p.3], Two-party computation 28 [p.1], Softmax 25 [p.2], Matrix multiplication 22 [p.1], Semi-honest model 8 [p.4], LayerNorm 7 [p.4], Fine-tuning 4 [p.11], Three-party computation 4 [p.11], Taylor expansion 3 [p.10], Rotations 3 [p.5], ReLU 2 [p.1], Sigmoid 2 [p.2], Attention 2 [p.4], Quantization 2 [p.14]
  • assumption: RLWE 35 [p.3], LWE 3 [p.14], Lattice-based 3 [p.3]
  • task: Language modeling / generation 2 [p.3]
  • topic: Homomorphic encryption 39 [p.1], Private / secure inference 18 [p.1], Open source 8 [p.3], Benchmark 6 [p.1]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/catalogue/5-072_bumblebee-secure-two-party-inference-framework-for-large-tra.pdf catalogue 18 1.0 MB 15583a865939

Database key: bumblebee-secure-two-party-inference-framework-for-large-transformers · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set bumblebee-secure-two-party-inference-framework-for-large-transformers COLUMN VALUE