Skip to content

CrypTFlow2: Practical 2-Party Secure Inference

Authors: Rathee, Rathee, Kumar, Chandran, Gupta, Rastogi, Sharma
Year: 2020
Venue: CCS 2020

type: system collection: catalogue status: auto

Links: ePrint 2020/1002

Summary

Auto-extracted summary (unreviewed)

CrypTFlow2 is a two-party secure inference framework built around new 2PC protocols for secure comparison and division that balance rounds and communication and produce outputs bitwise identical to cleartext fixed-point execution. It delivers the first secure inference of ImageNet-scale networks (ResNet50, DenseNet121, SqueezeNet) and, on prior benchmarks, needs an order of magnitude less communication and 20-30x less time than the state of the art, with ReLU layers 2-25x faster than garbled circuits.

Abstract

We present CrypTFlow2, a cryptographic framework for secure inference over realistic Deep Neural Networks (DNNs) using secure 2-party computation. CrypTFlow2 protocols are both correct – i.e., their outputs are bitwise equivalent to the cleartext execution – and efficient – they outperform the state-of-the-art protocols in both latency and scale. At the core of CrypTFlow2, we have new 2PC protocols for secure comparison and division, designed carefully to balance round and communication complexity for secure inference tasks. Using CrypTFlow2, we present the first secure inference over ImageNet-scale DNNs like ResNet50 and DenseNet121. These DNNs are at least an order of magnitude larger than those considered in the prior work of 2-party DNN inference. Even on the benchmarks considered by prior work, CrypTFlow2 requires an order of magnitude less communication and 20×-30× less time than the state-of-the-art.
Source: pdf

Extracted details

Field Value Source Confidence
contribution new 2PC protocols for secure comparison (millionaires) and division llm 0.90
contribution correct secure fixed-point arithmetic with faithful truncation (bitwise equal to cleartext) llm 0.90
contribution first secure inference over ImageNet-scale DNNs (ResNet50, DenseNet121) llm 0.90
contribution protocols over both Z_{2^l} and Z_n, implemented in the EzPC framework llm 0.90
schemes secret sharing llm 0.90
schemes HE llm 0.90
libraries EzPC llm 0.90
libraries SEAL llm 0.90
libraries EMP-toolkit llm 0.90
models SqueezeNet llm 0.90
models ResNet-50 llm 0.90
models DenseNet-121 llm 0.90
models MiniONN CNN llm 0.90
models ResNet-32 llm 0.90
datasets ImageNet llm 0.90
datasets CIFAR-10 llm 0.90
datasets CIFAR-100 llm 0.90
tasks image classification llm 0.90
tasks secure two-party DNN inference llm 0.90
hardware 3.7 GHz Intel Xeon, 4 cores, 16 GB RAM llm 0.90
hardware LAN 377 MBps / 0.3 ms llm 0.90
hardware WAN 40 MBps / 80 ms llm 0.90
security level 128-bit llm 0.90
threat model two-party client-server (2PC) llm 0.90
interaction interactive (HE+MPC) llm 0.90
code url https://github.com/mpc-msri/EzPC.1 (URL found in the full text) regex 0.50
code url https://github.com/mpc-msri/EzPC (URL found in the full text) regex 0.50
code url https://github.com/Microsoft/SEAL.Microsoft (URL found in the full text) regex 0.50
code url https://github.com/mpc-msri/EzPC llm 0.90
llm relevance medium llm 0.90
key result time and communication vs state of the art: 20-30x less time, ~10x less communication (benchmarks of prior work) llm 0.90
key result ReLU layer runtime: 16.4 s LAN / 69.4 s WAN / 5.23 GiB (ResNet50 over Z_{2^l}, vs 136.5 s / 1285.2 s / 39.19 GiB with garbled circuits) llm 0.90
key result ReLU speed-up over garbled circuits: 2-25x (up to 8x LAN, 18x WAN over Z_{2^l}) (SqueezeNet, ResNet50, DenseNet121) llm 0.90
key result non-linear layer improvement vs Delphi: 30.2x (MiniONN), 22.0x (ResNet32) (LAN, 41-bit) llm 0.90
key result online runtime improvement vs Delphi: 12.4x (MiniONN), 11.09x (ResNet32) (LAN) llm 0.90
key result Avgpool improvement over garbled circuits: up to 51x runtime, 41x communication (benchmark Avgpool layers) llm 0.90

System card: CrypTFlow2

Group 5. Interactive MPC and HE+MPC hybrids
Family OT + HE 2PC
Arithmetic Exact fixed point
Interaction Interactive 2PC
Model scale ResNet-50, DenseNet
Status Foundational
Relation to the SoK Exact 2PC comparison and truncation protocols.
Note Part of Microsoft EzPC. Basis for SIRNN and later LLM protocols.
Code https://github.com/mpc-msri/EzPC
Systems page CrypTFlow2

Keyword mentions

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

  • scheme: Garbled circuits 18 [p.2], MPC 15 [p.1], Secret sharing 9 [p.3], Oblivious transfer 8 [p.1], BFV 2 [p.4]
  • library: EzPC / CrypTFlow 24 [p.1], TensorFlow 8 [p.10], Microsoft SEAL 5 [p.10], nGraph-HE 2 [p.13], ABY / ABY3 2 [p.13], EVA / CHET 1 [p.13], MP-SPDZ 1 [p.13]
  • model: ResNet 25 [p.1], CNN 2 [p.2], AlexNet 1 [p.14]
  • dataset: ImageNet 15 [p.1], MNIST 2 [p.2], CIFAR-10 1 [p.12], CIFAR-100 1 [p.12], Chest X-ray 1 [p.2]
  • hardware: Intel Xeon 1 [p.11]
  • technique: ReLU 44 [p.1], Convolution 34 [p.2], Rotations 14 [p.10], Two-party computation 11 [p.1], Matrix multiplication 7 [p.2], Semi-honest model 5 [p.1], Quantization 3 [p.13], Three-party computation 2 [p.10], Malicious model 2 [p.4], Compiler 2 [p.13]
  • task: Medical / healthcare 2 [p.1]
  • topic: Benchmark 28 [p.1], Private / secure inference 23 [p.1], Homomorphic encryption 18 [p.2], Open source 6 [p.1]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/catalogue/5-063_cryptflow2-practical-2-party-secure-inference.pdf catalogue 18 916.0 KB 33a5042e1e11

Database key: cryptflow2-practical-2-party-secure-inference · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set cryptflow2-practical-2-party-secure-inference COLUMN VALUE