Skip to content

Cheetah: Lean and Fast Secure Two-Party Deep Neural Network Inference

Authors: Huang, Lu, Hong, Ding
Year: 2022
Venue: USENIX Security 2022

type: system collection: catalogue status: auto

Links: ePrint 2022/207 · Semantic Scholar

Summary

Auto-extracted summary (unreviewed)

Cheetah is a secure two-party DNN inference system that co-designs lattice-based HE, oblivious transfer and secret sharing: rotation-free HE protocols evaluate convolution, batch normalization and fully-connected layers, while lean OT-based primitives handle ReLU, truncation and comparison. A ResNet50 inference under a WAN takes under 2.5 minutes and 2.3 GB of communication, about 5.6x faster and 12.9x less communication than CrypTFlow2.

Abstract

Secure two-party neural network inference (2PC-NN) can offer privacy protection for both the client and the server and is a promising technique in the machine-learning-as-a-service setting. However, the large overhead of the current 2PC-NN inference systems is still being a headache, especially when applied to deep neural networks such as ResNet50. In this work, we present Cheetah, a new 2PC-NN inference system that is faster and more communication-efficient than state-of-the-arts. The main contributions of Cheetah are two-fold: the first part includes carefully designed homomorphic encryption-based protocols that can evaluate the linear layers (namely convolution, batch normalization, and fully-connection) without any expensive rotation operation. The second part includes several lean and communication-efficient primitives for the non-linear functions (e.g., ReLU and truncation). Using Cheetah, we present intensive benchmarks over several large-scale deep neural networks. Take ResNet50 for an example, an endto-end execution of Cheetah under a WAN setting costs less than 2.5 minutes and 2.3 gigabytes of communication, which outperforms CrypTFlow2 (ACM CCS 2020) by about 5.6× and 12.9×, respectively.
Source: pdf

Extracted details

Field Value Source Confidence
contribution HE protocols for linear layers (conv, BN, FC) without homomorphic rotations llm 0.90
contribution communication-efficient primitives for non-linear functions (millionaire, truncation, ReLU) llm 0.90
contribution linear protocols accept additive shares over Z_2^l, friendlier for subsequent non-linear layers llm 0.90
contribution extensive benchmarks on large DNNs including ResNet50 llm 0.90
contribution open-source implementation (OpenCheetah) llm 0.90
schemes lattice-based HE (RLWE, SEAL) llm 0.90
schemes oblivious transfer llm 0.90
schemes secret sharing llm 0.90
schemes VOLE-style OT extension (Ferret) llm 0.90
libraries SEAL llm 0.90
libraries HEXL llm 0.90
libraries EMP toolkit llm 0.90
libraries SCI (CrypTFlow2) llm 0.90
models ResNet-50 llm 0.90
models DenseNet llm 0.90
datasets ImageNet llm 0.90
tasks private DNN inference llm 0.90
tasks image classification llm 0.90
hardware Alibaba Cloud ecs.c7.2xlarge instances (2.70 GHz CPU, 16 GB RAM) llm 0.90
security level 128-bit llm 0.90
threat model two-party client-server (2PC), server holds model and client holds input llm 0.90
interaction interactive (HE+MPC) llm 0.90
code url https://github.com/Alibaba-Gemini-Lab/OpenCheetah (URL found in the full text) regex 0.50
code url https://github.com/EliSchwartz/imagenet-sample-images (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/Alibaba-Gemini-Lab/OpenCheetah llm 0.90
limitations orthogonal optimizations such as quantization and hardware acceleration are left as future work llm 0.90
llm relevance medium llm 0.90
key result end-to-end latency: < 2.5 minutes (ResNet50, WAN (44 MBps, 40 ms RTT)) llm 0.90
key result communication: 2.3 GB (ResNet50, WAN) llm 0.90
key result improvement over CrypTFlow2: 5.6x faster, 12.9x less communication (ResNet50, WAN) llm 0.90
key result linear-layer speed-up: 1.3x to 20x faster, 1.5x to 2x less communication (vs SCI-HE, single thread) llm 0.90
key result truncation with known MSB: 6.8x (LAN) / 7.1x (WAN) faster, 88x less communication (2^16 protocol runs vs CrypTFlow2) llm 0.90

System card: Cheetah

Group 5. Interactive MPC and HE+MPC hybrids
Family RLWE (BFV) + silent OT
Arithmetic Exact fixed point
Interaction Interactive 2PC
Model scale ResNet-50
Status Foundational
Relation to the SoK Coefficient-encoded HE matmul without rotations plus OT non-linearities.
Note Basis of Iron, BOLT and BumbleBee.
Code https://github.com/Alibaba-Gemini-Lab/OpenCheetah
Systems page Cheetah

Keyword mentions

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

  • scheme: Secret sharing 8 [p.1], Oblivious transfer 8 [p.1], MPC 5 [p.3]
  • library: EzPC / CrypTFlow 22 [p.1], Microsoft SEAL 9 [p.12], nGraph-HE 2 [p.14], MP-SPDZ 2 [p.16], ABY / ABY3 2 [p.15], HElib 1 [p.15], Concrete (Zama) 1 [p.12], EVA / CHET 1 [p.15]
  • model: ResNet 19 [p.1], RNN / LSTM 1 [p.16]
  • dataset: ImageNet 2 [p.1], MNIST 1 [p.1], Chest X-ray 1 [p.16]
  • technique: Two-party computation 41 [p.1], Convolution 25 [p.2], SIMD packing 17 [p.2], Rotations 16 [p.1], Semi-honest model 10 [p.6], ReLU 8 [p.1], Three-party computation 3 [p.13], Matrix multiplication 2 [p.7], Quantization 2 [p.14], Compiler 2 [p.14]
  • assumption: LWE 44 [p.3], RLWE 30 [p.3], Lattice-based 8 [p.1]
  • task: Language modeling / generation 2 [p.15]
  • topic: Homomorphic encryption 31 [p.1], Benchmark 9 [p.1], Private / secure inference 7 [p.1], Open source 7 [p.1]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/catalogue/5-064_cheetah-lean-and-fast-secure-two-party-deep-neural-network-i.pdf catalogue 20 426.8 KB bd5bf82f3ed9

Database key: cheetah-lean-and-fast-secure-two-party-deep-neural-network-inference · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set cheetah-lean-and-fast-secure-two-party-deep-neural-network-inference COLUMN VALUE