Skip to content

Orion: A Fully Homomorphic Encryption Framework for Deep Learning

Authors: Austin Ebel; Karthik Garimella; Brandon Reagen
Year: 2025
Venue: ASPLOS 2025

type: library collection: catalogue, sok-references status: auto

Links: DOI · arXiv:2311.03470

Summary

Auto-extracted summary (unreviewed)

Orion is a fully automated framework that takes PyTorch neural networks and compiles them into CKKS FHE programs through custom Lattigo bindings, using a single-shot multiplexed packing strategy for arbitrary convolutions plus automatic bootstrap placement, level and scale management and range estimation. It beats the state of the art by 2.38x on ResNet-20 (single-threaded), scales to ResNet-50 on ImageNet, and shows the first high-resolution FHE object detection with a 139M-parameter YOLO-v1.

Abstract

Fully Homomorphic Encryption (FHE) has the potential to substantially improve privacy and security by enabling computation directly on encrypted data. This is especially true with deep learning, as today, many popular user services are powered by neural networks in the cloud. Beyond its well-known high computational costs, one of the major challenges facing wide-scale deployment of FHE-secured neural inference is effectively mapping these networks to FHE primitives. FHE poses many programming challenges including packing large vectors, managing accumulated noise, and translating arbitrary and general-purpose programs to the limited instruction set provided by FHE. These challenges make building large FHE neural networks intractable using the tools available today. In this paper we address these challenges with Orion, a fully-automated framework for private neural inference using FHE. Orion accepts deep neural networks written in PyTorch and translates them into efficient FHE programs. We achieve this by proposing a novel single-shot multiplexed packing strategy for arbitrary convolutions and through a new, efficient technique to automate bootstrap placement and scale management. We evaluate Orion on common benchmarks used by the FHE deep learning community and outperform state-of-the-art by 2.38x on ResNet-20, the largest network they report. Orion's techniques enable processing much deeper and larger networks. We demonstrate this by evaluating ResNet-50 on ImageNet and present the first high-resolution FHE object detection experiments using a YOLO-v1 model with 139 million parameters. Orion is open-source for all to use at: https://github.com/baahl-nyu/orion
Source: arxiv

Extracted details

Field Value Source Confidence
contribution automated PyTorch-to-FHE framework with high-level orion.nn API llm 0.90
contribution single-shot multiplexed packing for arbitrary convolutions llm 0.90
contribution automatic bootstrap placement and level management llm 0.90
contribution error-free scaling-factor management llm 0.90
contribution automatic range estimation and Chebyshev activation fitting llm 0.90
contribution open-source release llm 0.90
schemes CKKS llm 0.90
libraries Lattigo llm 0.90
libraries PyTorch llm 0.90
models ResNet-20 llm 0.90
models ResNet-50 llm 0.90
models YOLO-v1 llm 0.90
models AlexNet llm 0.90
models ResNet-1202 llm 0.90
datasets ImageNet llm 0.90
tasks image classification llm 0.90
tasks object detection llm 0.90
tasks private neural network inference llm 0.90
security level 128-bit llm 0.90
threat model outsourced cloud inference on client-encrypted data llm 0.90
interaction non-interactive FHE llm 0.90
code url https://github.com/baahl-nyu/orion.Permission (URL found in the full text) regex 0.50
code url https://github.com/tuneinsight/lattigo (URL found in the full text) regex 0.50
code url http://github.com/jax-ml/jax (URL found in the full text) regex 0.50
code url https://github.com/baahl-nyu/orion llm 0.90
limitations single-threaded Lattigo backend; multi-threaded and GPU backends left for future work llm 0.90
limitations self-attention layer support not yet included llm 0.90
llm relevance medium llm 0.90
key result speed-up over state of the art: 2.38x (ResNet-20, single-threaded) llm 0.90
key result bootstrap placement time: 1.94 s (8.14x faster than Dacapo) (ResNet-20) llm 0.90
key result bootstrap placement time: 0.82 s (1270x faster than Dacapo) (AlexNet) llm 0.90
key result bootstrap placement time: 151 s (ResNet-1202) llm 0.90
key result largest network evaluated: YOLO-v1 with 139M parameters on 448x448x3 images (FHE object detection) llm 0.90

System card: Orion

Group 7. FHE libraries, GPU back-ends and compilers
Family CKKS (Lattigo back-end)
Arithmetic Python/Go
Interaction Compiler/framework
Model scale CACHEMIR extends its bootstrapping placement
Status Active SOTA
Relation to the SoK PyTorch-to-CKKS compiler with automatic bootstrapping placement.
Note Shortest-path bootstrapping placement reused by CACHEMIR.
Code https://github.com/baahl-nyu/orion
Systems page Orion

SoK reference [85]

A. Ebel, K. Garimella, and B. Reagen, “Orion: A fully homomorphic encryption framework for deep learning,” in Proceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2, 2025, pp. 734–749.

Listed on the reference page of the site.

Keyword mentions

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

  • scheme: CKKS 36 [p.1]
  • library: Orion 90 [p.1], PyTorch 18 [p.1], Fhelipe 14 [p.2], EVA / CHET 10 [p.10], HEaaN (library) 5 [p.10], Microsoft SEAL 4 [p.13], OpenFHE 3 [p.10], Lattigo 3 [p.9], HELayers 3 [p.12], nGraph-HE 3 [p.13], TenSEAL 2 [p.13], Cheddar 2 [p.13], HElib 1 [p.15]
  • model: ResNet 48 [p.1], MobileNet 9 [p.10], AlexNet 6 [p.8], CNN 6 [p.2], VGG 4 [p.10], LeNet 3 [p.10], MLP 3 [p.10]
  • dataset: ImageNet 20 [p.1], MNIST 9 [p.10], CIFAR-10 9 [p.10]
  • hardware: GPU 5 [p.10], Intel Xeon 1 [p.10], CUDA 1 [p.15]
  • technique: Convolution 52 [p.1], Rotations 39 [p.2], Bootstrapping 26 [p.2], ReLU 17 [p.7], Rescaling 12 [p.2], SIMD packing 12 [p.2], Compiler 12 [p.13], RNS 8 [p.3], Key switching 4 [p.4], Chebyshev 4 [p.9], Minimax approximation 3 [p.10], NTT 2 [p.4], Fine-tuning 2 [p.12]
  • assumption: 128-bit security 3 [p.3]
  • task: Object detection 8 [p.1], Language modeling / generation 2 [p.3]
  • topic: Homomorphic encryption 71 [p.1], Open source 8 [p.1], Benchmark 6 [p.2], Private / secure inference 4 [p.10], Privacy-preserving machine learning 2 [p.1]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/catalogue/7-104_orion-a-fully-homomorphic-encryption-framework-for-deep-lear.pdf catalogue 16 3.1 MB dcff5492c59d
references/sok-references/085_orion-a-fully-homomorphic-encryption-framework-for-deep-lear.pdf sok-references 16 3.1 MB dcff5492c59d byte-identical to references/catalogue/7-104_orion-a-fully-homomorphic-encryption-framework-for-deep-lear.pdf

Database key: orion-a-fully-homomorphic-encryption-framework-for-deep-learning · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set orion-a-fully-homomorphic-encryption-framework-for-deep-learning COLUMN VALUE