Skip to content

Delphi: A Cryptographic Inference Service for Neural Networks

Authors: Pratyush Mishra; Ryan Lehmkuhl; Akshayaram Srinivasan; Wenting Zheng; Raluca Ada Popa
Year: 2020
Venue: USENIX Security 2020

type: system collection: catalogue status: auto

Links: DOI · ePrint 2020/50 · link · Semantic Scholar

Summary

Auto-extracted summary (unreviewed)

Delphi is a two-party cryptographic inference system for convolutional networks that co-designs cryptography and ML. A hybrid protocol moves the heavy homomorphic-encryption work for linear layers into an input-independent preprocessing phase, leaving a cheap secret-sharing online phase (with GPU-friendly prime-field convolutions) plus garbled circuits for ReLU, and a planner selects which ReLUs to replace with cheap quadratic activations. On CIFAR-10 (MiniONN CNN) and CIFAR-100 (ResNet-32) this yields a 22x reduction in online prediction latency versus GAZELLE.

Abstract

Many companies provide neural network prediction services to users for a wide range of applications. However, current prediction systems compromise one party’s privacy: either the user has to send sensitive inputs to the service provider for classification, or the service provider must store its proprietary neural networks on the user’s device. The former harms the personal privacy of the user, while the latter reveals the service provider’s proprietary model. We design, implement, and evaluate DELPHI, a secure prediction system that allows two parties to execute neural network inference without revealing either party’s data. DELPHI approaches the problem by simultaneously co-designing cryptography and machine learning. We first design a hybrid cryptographic protocol that improves upon the communication and computation costs over prior work. Second, we develop a planner that automatically generates neural network architecture configurations that navigate the performance-accuracy trade-offs of our hybrid protocol. Together, these techniques allow us to achieve a 22× improvement in online prediction latency compared to the state-of-the-art prior work.
Source: pdf

Extracted details

Field Value Source Confidence
contribution hybrid cryptographic protocol with input-independent preprocessing that cuts online computation and communication llm 0.90
contribution GPU-accelerated online linear layers enabled by the choice of prime field llm 0.90
contribution planner that automatically searches network configurations mixing ReLU and quadratic activations llm 0.90
contribution first secure inference evaluation on CIFAR-100 with ResNet-32 llm 0.90
schemes HE llm 0.90
schemes secret sharing llm 0.90
schemes garbled circuits llm 0.90
schemes Beaver triples llm 0.90
libraries SEAL llm 0.90
libraries fancy-garbling llm 0.90
models MiniONN 7-layer CNN llm 0.90
models ResNet-32 llm 0.90
datasets CIFAR-10 llm 0.90
datasets CIFAR-100 llm 0.90
tasks image classification llm 0.90
tasks cryptographic neural network inference llm 0.90
hardware AWS c5.2xlarge (Intel Xeon 8000 series 3.0GHz, 16GB RAM), client and server in different regions llm 0.90
hardware NVIDIA Tesla V100 GPUs (ML experiments) llm 0.90
threat model two-party client-server llm 0.90
interaction interactive (HE+MPC) llm 0.90
code url https://github.com/GaloisInc/fancy-garbling/Planner (URL found in the full text) regex 0.50
code url https://github.com/chiraag/gazelle_mpc (URL found in the full text) regex 0.50
code url https://github.com/mc2-project/delphi llm 0.90
limitations preprocessing time and communication are higher than GAZELLE's (at most equal to GAZELLE's online cost) llm 0.90
limitations GPU use requires copying weights and inputs to and from GPU memory for every linear layer llm 0.90
llm relevance medium llm 0.90
key result online prediction latency improvement: 22x (vs state-of-the-art prior work (GAZELLE)) llm 0.90
key result online convolution time: over 80x smaller than GAZELLE (ResNet-32 convolutions) llm 0.90
key result online convolution communication: over 150x lower than GAZELLE (ResNet-32 convolutions) llm 0.90
key result GPU convolution cost: 50-100x lower than CPU (single convolutions, batch sizes 1/5/10) llm 0.90

System card: Delphi

Group 5. Interactive MPC and HE+MPC hybrids
Family HE preprocessing + GC
Arithmetic Exact integers
Interaction Interactive 2PC
Model scale CNNs
Status Foundational
Relation to the SoK Moves HE to an offline preprocessing phase.
Note Fast online phase. Also co-designs activations with polynomial approximation.
Code https://github.com/mc2-project/delphi
Systems page Delphi

Keyword mentions

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

  • scheme: Garbled circuits 26 [p.2], Secret sharing 17 [p.1], TEE / confidential computing 13 [p.12], Oblivious transfer 8 [p.5], Differential privacy 4 [p.3], MPC 3 [p.1], Homomorphic secret sharing 3 [p.12], Paillier 2 [p.15]
  • library: Microsoft SEAL 5 [p.9], EVA / CHET 3 [p.13], ABY / ABY3 2 [p.12], EzPC / CrypTFlow 2 [p.13], HElib 1 [p.16], MP-SPDZ 1 [p.13]
  • model: ResNet 20 [p.1], CNN 2 [p.2], SVM 2 [p.12], Linear regression 2 [p.12], Logistic regression 1 [p.14], Decision tree 1 [p.17], k-NN 1 [p.17]
  • dataset: CIFAR-10 8 [p.10], CIFAR-100 7 [p.2]
  • hardware: GPU 7 [p.2], NVIDIA V100 1 [p.10], Intel Xeon 1 [p.9], ARM 1 [p.14]
  • technique: ReLU 53 [p.2], Convolution 17 [p.1], Two-party computation 13 [p.3], Polynomial approximation 5 [p.5], Semi-honest model 4 [p.3], Pruning 2 [p.16]
  • assumption: LWE 2 [p.15], Lattice-based 2 [p.15]
  • topic: Homomorphic encryption 56 [p.1], Private / secure inference 8 [p.10], Open source 5 [p.9]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/catalogue/5-062_delphi-a-cryptographic-inference-service-for-neural-networks.pdf catalogue 19 5.8 MB 1b72ff554f73

Database key: delphi-a-cryptographic-inference-service-for-neural-networks · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set delphi-a-cryptographic-inference-service-for-neural-networks COLUMN VALUE