Skip to content

BOLT: Privacy-Preserving, Accurate and Efficient Inference for Transformers

Authors: Pang, Zhu, Lou, Wang
Year: 2024
Venue: IEEE S&P 2024

type: system collection: catalogue status: auto

Links: ePrint 2023/1893

Summary

Auto-extracted summary (unreviewed)

BOLT is a two-party privacy-preserving inference framework for transformers that combines HE (SEAL) for matrix multiplications with secret-sharing protocols (SCI/EzPC) for GELU, Softmax and LayerNorm, together with ML-side optimizations such as secure-computation-aware fine-tuning and word elimination. It cuts communication by 10.91x and runs 4.8-9.5x faster than the prior state of the art (Iron) across LAN/WAN settings while matching plaintext BERT accuracy on GLUE tasks.

Abstract

The advent of transformers has brought about significant advancements in traditional machine learning tasks. However, their pervasive deployment has raised concerns about the potential leakage of sensitive information during inference. Existing approaches using secure multiparty computation (MPC) face limitations when applied to transformers due to the extensive model size and resource-intensive matrix-matrix multiplications. In this paper, we present BOLT, a privacypreserving inference framework for transformer models that supports efficient matrix multiplications and nonlinear computations. Combined with our novel machine learning optimizations, BOLT reduces the communication cost by 10.91×. Our evaluation on diverse datasets demonstrates that BOLT maintains comparable accuracy to floating-point models and achieves 4.8-9.5× faster inference across various network settings compared to the state-of-the-art system.
Source: pdf

Extracted details

Field Value Source Confidence
contribution efficient HE protocols for ciphertext-plaintext and ciphertext-ciphertext matrix multiplication in transformers llm 0.90
contribution optimized secure protocols for nonlinear functions (GELU, Softmax, LayerNorm) llm 0.90
contribution secure computation-aware fine-tuning and word elimination to reduce cost llm 0.90
contribution 10.91x communication reduction and 4.8-9.5x speed-up over Iron llm 0.90
contribution open-source implementation and faithful re-implementation of Iron for comparison llm 0.90
schemes BFV llm 0.90
schemes secret sharing llm 0.90
libraries SCI (EzPC) llm 0.90
libraries SEAL llm 0.90
models BERT llm 0.90
datasets MRPC llm 0.90
datasets RTE llm 0.90
datasets SST-2 llm 0.90
datasets STS-B llm 0.90
tasks private transformer inference llm 0.90
tasks text classification llm 0.90
tasks sentence similarity regression llm 0.90
hardware AWS EC2 c6i.16xlarge (64 vCPUs, 128 GB RAM), 32 threads llm 0.90
security level 128-bit llm 0.90
threat model semi-honest two-party client-server; circuit privacy via noise flooding llm 0.90
interaction interactive (HE+MPC) llm 0.90
code url https://github.com/Clive2312/BOLT.2 (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/Clive2312/BOLT llm 0.90
limitations conversion from MPC back to HE is omitted, relying on shared values being small relative to the ring size to avoid overflow llm 0.90
llm relevance high llm 0.90
key result communication reduction: 10.91x (vs state-of-the-art system (Iron)) llm 0.90
key result end-to-end inference speed-up: 4.8-9.5x (vs Iron across LAN (3 Gbps, 0.8 ms) and WAN (100/200 Mbps, 40/80 ms) settings) llm 0.90
key result SST-2 accuracy: 92.78 (BOLT with word elimination) vs 92.36 plaintext (GLUE validation set) llm 0.90
key result MRPC F1: 89.95 (BOLT with word elimination) vs 90.00 plaintext (GLUE validation set) llm 0.90
key result STS-B Pearson correlation: 88.44 (BOLT with word elimination) vs 89.62 plaintext (GLUE validation set) llm 0.90

System card: BOLT

Group 5. Interactive MPC and HE+MPC hybrids
Family HE (BFV) + OT 2PC
Arithmetic Fixed point
Interaction Interactive 2PC
Model scale BERT-Base
Status Active SOTA
Relation to the SoK Hybrid with compact HE packing and word elimination.
Note About 10x less communication than Iron. Still interactive.
Code https://github.com/Clive2312/BOLT
Systems page BOLT

Keyword mentions

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

  • scheme: MPC 57 [p.1], Secret sharing 16 [p.2], BFV 4 [p.3], Differential privacy 2 [p.4]
  • library: EzPC / CrypTFlow 4 [p.11], Microsoft SEAL 3 [p.11], TensorFlow 1 [p.14]
  • model: BERT 45 [p.1], Transformer 10 [p.1], GPT 4 [p.14], CNN 3 [p.1], RNN / LSTM 1 [p.16]
  • dataset: GLUE tasks (QNLI/RTE/MRPC/CoLA/QQP/MNLI/STS-B) 11 [p.11], SST-2 4 [p.11], GLUE 1 [p.11], ECG 1 [p.15]
  • hardware: GPU 2 [p.14]
  • technique: Matrix multiplication 62 [p.1], Rotations 49 [p.2], Softmax 39 [p.2], GELU 33 [p.2], LayerNorm 19 [p.3], Fine-tuning 17 [p.2], Attention 12 [p.1], Quantization 10 [p.10], Two-party computation 9 [p.4], Rescaling 8 [p.13], Bootstrapping 6 [p.3], Malicious model 5 [p.14], Polynomial approximation 4 [p.8], Remez 4 [p.9], ReLU 4 [p.9], Lookup table 4 [p.9], Semi-honest model 4 [p.4], SIMD packing 3 [p.4], Convolution 3 [p.1], Relinearization 2 [p.11], Knowledge distillation 2 [p.14]
  • assumption: Lattice-based 3 [p.3]
  • task: Medical / healthcare 7 [p.1]
  • topic: Homomorphic encryption 78 [p.2], Private / secure inference 13 [p.2], Privacy-preserving machine learning 4 [p.4], Open source 3 [p.2], Benchmark 2 [p.11]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/catalogue/5-071_bolt-privacy-preserving-accurate-and-efficient-inference-for.pdf catalogue 19 1.4 MB 49d8119866f7

Database key: bolt-privacy-preserving-accurate-and-efficient-inference-for · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set bolt-privacy-preserving-accurate-and-efficient-inference-for COLUMN VALUE