Skip to content

HE-MAN – Homomorphically Encrypted MAchine learning with oNnx models

Authors: Martin Nocker; David Drexel; Michael Rader; Alessio Montuoro; Pascal Schöttle
Year: 2023
Venue: Proceedings of the 2023 8th International Conference on Machine Learning Technologies

type: library collection: ppml status: auto

Links: DOI

Summary

Auto-extracted summary (unreviewed)

HE-MAN is an open-source two-party toolset for privacy-preserving inference on ONNX models with homomorphically encrypted inputs, with backends on Concrete (HE-MAN-Concrete) and TenSEAL (HE-MAN-TenSEAL). It hides cryptographic details from both model owner and data owner: a calibration dataset and a target security level (128-bit by default) are used to derive encryption parameters automatically, the client generates keys and encrypts, and the server runs inference on the calibrated ONNX graph. On digit-classification and face-recognition networks the encrypted accuracy matches plaintext while latency is several orders of magnitude higher.

Abstract

Machine learning (ML) algorithms are increasingly important for the success of products and services, especially considering the growing amount and availability of data. This also holds for areas handling sensitive data, e.g. applications processing medical data or facial images. However, people are reluctant to pass their personal sensitive data to a ML service provider. At the same time, service providers have a strong interest in protecting their intellectual property and therefore refrain from publicly sharing their ML model. Fully homomorphic encryption (FHE) is a promising technique to enable individuals using ML services without giving up privacy and protecting the ML model of service providers at the same time. Despite steady improvements, FHE is still hardly integrated in today’s ML applications. Reasons for that are, among others, that existing implementations either require the user to possess expertise in FHE, do not feature an easy ML framework integration, or have to approximate non-polynomial activations. We introduce HE-MAN, an open-source two-party machine learning toolset for privacy preserving inference with ONNX models and homomorphically encrypted data. Both the model and the input data do not have to be disclosed. HE-MAN abstracts cryptographic details away from the users, thus expertise in FHE is not required for either party. HE-MAN’s security relies on its underlying FHE schemes. For now, we integrate two different homomorphic encryption schemes, namely Concrete and TenSEAL. Compared to prior work, HE-MAN supports a broad range of ML models in ONNX format out of the box without sacrificing accuracy. We evaluate the performance of our implementation on different network architectures classifying handwritten digits and performing face recognition and report accuracy and latency of the homomorphically encrypted inference. Cryptographic parameters ∗Correspondence: martin.nocker@mci.edu, MCI The Entrepreneurial School, Universitätsstraße 15, 6020 Innsbruck, Austria This work is licensed under a Creative Commons Attribution International 4.0 License. ICMLT 2023, March 10–12, 2023, Stockholm, Sweden © 2023 Copyright held by the owner/author(s). ACM ISBN 978-1-4503-9832-9/23/03. https://doi.org/10.1145/3589883.3589889 are automatically derived by the tools. We show that the accuracy of HE-MAN is on par with models using plaintext input while inference latency is several orders of magnitude higher compared to the plaintext case.
Source: pdf

Extracted details

Field Value Source Confidence
contribution open-source two-party FHE inference toolset accepting ONNX models llm 0.75
contribution automatic derivation of encryption parameters from the model, a calibration dataset and a security level llm 0.75
contribution two backends: Concrete and TenSEAL llm 0.75
contribution command-line workflow covering parameter derivation, key generation, encryption, inference and decryption llm 0.75
contribution evaluation of accuracy and latency on digit classification and face recognition llm 0.75
schemes TFHE llm 0.75
schemes CKKS llm 0.75
libraries Concrete llm 0.75
libraries TenSEAL llm 0.75
libraries ONNX llm 0.75
libraries PyTorch llm 0.75
models CNN (handwritten digit classification) llm 0.75
models face recognition network llm 0.75
datasets handwritten digit dataset (unnamed in excerpt) llm 0.75
datasets face recognition dataset (unnamed in excerpt) llm 0.75
tasks handwritten digit classification llm 0.75
tasks face recognition llm 0.75
tasks privacy-preserving ML inference llm 0.75
security level 128-bit by default (80-bit Concrete parameters also listed) llm 0.75
threat model two-party: model owner (server) and data owner (client); both input data and model kept private llm 0.75
interaction non-interactive FHE llm 0.75
code url https://github.com/Microsoft/SEAL.Microsoft (URL found in the full text) regex 0.50
code url https://github.com/smile-ffg/he-man-concrete (URL found in the full text) regex 0.50
code url https://github.com/smile-ffg/he-man-tenseal (URL found in the full text) regex 0.50
code url https://github.com/smile-ffg/he-man-concrete llm 0.75
limitations inference latency is several orders of magnitude higher than plaintext llm 0.75
limitations execution time depends heavily on the underlying open-source FHE libraries llm 0.75
limitations calibration dataset must be representative of the inference task to derive proper parameters llm 0.75
llm relevance medium llm 0.75
key result accuracy: on par with plaintext models (digit classification and face recognition under HE) llm 0.75
key result inference latency: several orders of magnitude higher than plaintext (encrypted inference with Concrete / TenSEAL backends) llm 0.75

Keyword mentions

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

  • scheme: CKKS 10 [p.3], TFHE 6 [p.3], Paillier 4 [p.3], BFV 2 [p.3]
  • library: TenSEAL 32 [p.2], ONNX 29 [p.2], Concrete (Zama) 28 [p.2], Microsoft SEAL 9 [p.3], EVA / CHET 8 [p.9], nGraph-HE 3 [p.9], OpenFHE 2 [p.3], HElib 2 [p.3], HELayers 2 [p.9], CrypTen 2 [p.9], EzPC / CrypTFlow 2 [p.9], TensorFlow 2 [p.9], Lattigo 1 [p.10], PyTorch 1 [p.9]
  • model: LeNet 14 [p.7], ResNet 2 [p.6], CNN 1 [p.10], Logistic regression 1 [p.10]
  • dataset: MNIST 9 [p.7]
  • hardware: GPU 1 [p.9]
  • technique: ReLU 27 [p.3], Bootstrapping 12 [p.3], Compiler 8 [p.9], Convolution 7 [p.3], Programmable bootstrapping 4 [p.3], Two-party computation 4 [p.2], Matrix multiplication 3 [p.6], Semi-honest model 3 [p.4], Polynomial approximation 2 [p.7], Square activation 2 [p.7], Encrypted training 2 [p.10], Model inversion 2 [p.4]
  • assumption: RLWE 6 [p.3], LWE 4 [p.3], Lattice-based 2 [p.10]
  • topic: Homomorphic encryption 159 [p.1], Open source 8 [p.2], Private / secure inference 6 [p.2]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/ppml/inference/nocker-2023_he-man-homomorphically-encrypted-machine-learning-with-onnx-models.pdf ppml/inference 12 714.4 KB c18114ef6064 was Nocker et al. - 2023 - HE-MAN – Homomorphically Encrypted MAchine learning with oNnx models.pdf

Database key: he-man-homomorphically-encrypted-machine-learning-with-onnx-models · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set he-man-homomorphically-encrypted-machine-learning-with-onnx-models COLUMN VALUE