Skip to content

Slalom: Fast, Verifiable and Private Execution of Neural Networks in Trusted Hardware

Authors: Florian Tramèr; Dan Boneh
Year: 2019
Venue: ICLR 2019

type: system collection: catalogue status: auto

Links: arXiv:1806.03287

Summary

Auto-extracted summary (unreviewed)

Slalom speeds up DNN inference inside a TEE by securely delegating all linear layers from the enclave (Intel SGX) to a faster but untrusted co-located GPU, building on an efficient outsourcing scheme for matrix multiplication that provides integrity and, optionally, privacy. For VGG16, MobileNet and ResNet variants it achieves 6-20x higher throughput for verifiable inference and 4-11x for verifiable and private inference compared with running the whole network in the enclave.

Abstract

As Machine Learning (ML) gets applied to security-critical or sensitive domains, there is a growing need for integrity and privacy for outsourced ML computations. A pragmatic solution comes from Trusted Execution Environments (TEEs), which use hardware and software protections to isolate sensitive computations from the untrusted software stack. However, these isolation guarantees come at a price in performance, compared to untrusted alternatives. This paper initiates the study of high performance execution of Deep Neural Networks (DNNs) in TEEs by efficiently partitioning DNN computations between trusted and untrusted devices. Building upon an efficient outsourcing scheme for matrix multiplication, we propose Slalom, a framework that securely delegates execution of all linear layers in a DNN from a TEE (e.g., Intel SGX or Sanctum) to a faster, yet untrusted, co-located processor. We evaluate Slalom by running DNNs in an Intel SGX enclave, which selectively delegates work to an untrusted GPU. For canonical DNNs (VGG16, MobileNet and ResNet variants) we obtain 6x to 20x increases in throughput for verifiable inference, and 4x to 11x for verifiable and private inference.
Source: arxiv

Extracted details

Field Value Source Confidence
contribution Framework partitioning DNN execution between a TEE and an untrusted co-located processor llm 0.85
contribution Secure outsourcing of all linear layers from the enclave to an untrusted GPU with verifiable results llm 0.85
contribution Optional privacy for the outsourced computation llm 0.85
contribution Evaluation with Intel SGX enclave delegating to an untrusted GPU on canonical DNNs llm 0.85
schemes TEE llm 0.85
models VGG-16 llm 0.85
models MobileNet llm 0.85
models ResNet llm 0.85
tasks verifiable and private DNN inference llm 0.85
tasks outsourced ML computation llm 0.85
hardware Intel SGX enclave llm 0.85
hardware GPU llm 0.85
hardware CPU with AVX llm 0.85
threat model client trusts the TEE; co-located GPU/processor is untrusted; integrity and privacy of outsourced ML computation llm 0.85
interaction TEE llm 0.85
code url https://github.com/ftramer/slalom.We (URL found in the full text) regex 0.50
code url https://github.com/ftramer/slalom.Slalom (URL found in the full text) regex 0.50
code url https://github.com/ftramer/slalom llm 0.85
llm relevance medium llm 0.85
key result throughput increase: 6x-20x (verifiable inference, VGG16/MobileNet/ResNet, SGX enclave + untrusted GPU) llm 0.85
key result throughput increase: 4x-11x (verifiable and private inference) llm 0.85

System card: Slalom

Group 6. Trusted hardware, statistical and integrity-only alternatives
Family SGX TEE + outsourced GPU linear layers
Arithmetic Native / field arithmetic
Interaction Trusted hardware
Model scale VGG, MobileNet
Status Foundational
Relation to the SoK Hybrid TEE plus untrusted GPU with blinding.
Note Template for TEE-GPU split designs.
Code https://github.com/ftramer/slalom
Systems page Slalom

Keyword mentions

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

  • scheme: TEE / confidential computing 136 [p.1]
  • library: TensorFlow 3 [p.7]
  • model: MobileNet 27 [p.1], VGG 26 [p.1], ResNet 18 [p.1], CNN 1 [p.11]
  • dataset: ImageNet 4 [p.7], CIFAR-10 1 [p.15]
  • hardware: GPU 17 [p.1], ASIC 3 [p.2], ARM 3 [p.1]
  • technique: Convolution 49 [p.4], Quantization 14 [p.4], Matrix multiplication 4 [p.1]
  • topic: Benchmark 17 [p.4], Private / secure inference 4 [p.1], Open source 2 [p.2]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/catalogue/6-085_slalom-fast-verifiable-and-private-execution-of-neural-netwo.pdf catalogue 19 669.1 KB ffdccc420574

Database key: slalom-fast-verifiable-and-private-execution-of-neural-networks-in · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set slalom-fast-verifiable-and-private-execution-of-neural-networks-in COLUMN VALUE