Skip to content

HeLayers: A Tile Tensors Framework for Large Neural Networks on Encrypted Data

Authors: Ehud Aharoni; Allon Adir; Moran Baruch; Nir Drucker; Gilad Ezov; Ariel Farkash; Lev Greenberg; Ramy Masalha; Guy Moshkowich; Dov Murik; Hayim Shaul; Omri Soceanu
Year: 2023
Venue: PETS 2023

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

Links: DOI · arXiv:2011.01805

Summary

Auto-extracted summary (unreviewed)

HeLayers is an IBM framework that abstracts the SIMD packing decision for HE computations through 'tile tensor' data structures and a packing optimizer, and introduces a new algorithm for 2D convolutions on packed ciphertexts. Using CKKS in Microsoft SEAL at 128-bit security, the framework runs non-interactive inference over an HE-friendly AlexNet with large 224x224x3 inputs in a few minutes (three minutes in the abstract metadata, around five minutes in the paper text), several orders of magnitude faster than the prior best HE-only result of more than three hours at 80-bit security.

Abstract

Privacy-preserving solutions enable companies to offload confidential data to third-party services while fulfilling their government regulations. To accomplish this, they leverage various cryptographic techniques such as Homomorphic Encryption (HE), which allows performing computation on encrypted data. Most HE schemes work in a SIMD fashion, and the data packing method can dramatically affect the running time and memory costs. Finding a packing method that leads to an optimal performant implementation is a hard task. We present a simple and intuitive framework that abstracts the packing decision for the user. We explain its underlying data structures and optimizer, and propose a novel algorithm for performing 2D convolution operations. We used this framework to implement an HE-friendly version of AlexNet, which runs in three minutes, several orders of magnitude faster than other state-of-the-art solutions that only use HE.
Source: arxiv

Extracted details

Field Value Source Confidence
contribution tile tensor data structure abstracting packing choices for HE llm 0.80
contribution packing optimizer that selects a layout under security and memory constraints llm 0.80
contribution novel algorithm for homomorphic 2D convolution llm 0.80
contribution HE-friendly AlexNet inference with large inputs in minutes using CKKS only llm 0.80
schemes CKKS llm 0.80
libraries SEAL llm 0.80
libraries HeLayers llm 0.80
models AlexNet (HE-friendly) llm 0.80
tasks encrypted CNN inference llm 0.80
tasks homomorphic 2D convolution llm 0.80
security level 128-bit llm 0.80
threat model non-interactive outsourced HE inference; client encrypts data, cloud evaluates without decryption llm 0.80
interaction non-interactive FHE llm 0.80
code url https://github.com/Microsoft/SEAL (URL found in the full text) regex 0.50
code url https://github.com/kaizouman/tensorsandbox/tree/master/cifar10/models/squeeze (URL found in the full text) regex 0.50
code url https://github.com/IBM/fhe-toolkit-linux/tree/42af59512bbf9fa58602a1d54d3e2e28b37dfb01/ (URL found in the full text) regex 0.50
llm relevance medium llm 0.80
key result AlexNet encrypted inference time: about 3-5 minutes (abstract says three minutes, paper text says around five minutes) (HE-friendly AlexNet, 224x224x3 input, CKKS via SEAL, 128-bit security) llm 0.80
key result prior best HE-only AlexNet inference: more than 3 hours (REDsec, 80-bit security) llm 0.80

System card: IBM HELayers

Group 7. FHE libraries, GPU back-ends and compilers
Family CKKS, BGV (via SEAL/HElib/HEaaN)
Arithmetic C++/Python
Interaction ML-on-HE toolkit
Model scale Used by Polyformer
Status Active
Relation to the SoK Tile-tensor packing abstraction for neural networks.
Note Free community edition. Not open source.
Code https://ibm.github.io/helayers/
Systems page IBM HELayers

SoK reference [92]

E. Aharoni, A. Adir, M. Baruch, N. Drucker, G. Ezov, A. Farkash, L. Greenberg, R. Masalha, G. Moshkowich, D. Murik, H. Shaul, and O. Soceanu, “HeLayers: A tile tensors framework for large neural networks on encrypted data,” Proceedings on Privacy Enhancing Technologies, vol. 2023, no. 1, pp. 325–342, Jan. 2023.

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 6 [p.1], MPC 2 [p.2]
  • library: EVA / CHET 20 [p.2], HELayers 16 [p.1], Microsoft SEAL 6 [p.4], HElib 5 [p.4], TenSEAL 3 [p.12], nGraph-HE 2 [p.12], PyTorch 1 [p.11]
  • model: AlexNet 40 [p.1], CNN 5 [p.13], ResNet 4 [p.12], MobileNet 4 [p.2], VGG 3 [p.12], LeNet 1 [p.13], Logistic regression 1 [p.14]
  • dataset: COVID 7 [p.1], MNIST 3 [p.10], CIFAR-10 1 [p.13], ImageNet 1 [p.14]
  • hardware: Intel Xeon 1 [p.15], GPU 1 [p.14]
  • technique: Convolution 77 [p.1], Rotations 23 [p.2], SIMD packing 13 [p.1], Matrix multiplication 13 [p.2], Non-interactive 5 [p.1], ReLU 3 [p.2], Malicious model 3 [p.11], Bootstrapping 2 [p.2], Semi-honest model 2 [p.4]
  • topic: Homomorphic encryption 83 [p.1], Benchmark 10 [p.9], Private / secure inference 2 [p.14], Open source 2 [p.14]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/catalogue/7-096_helayers-a-tile-tensors-framework-for-large-neural-networks.pdf catalogue 18 1.9 MB ffc20bb27daa
references/sok-references/092_helayers-a-tile-tensors-framework-for-large-neural-networks.pdf sok-references 18 1.9 MB ffc20bb27daa byte-identical to references/catalogue/7-096_helayers-a-tile-tensors-framework-for-large-neural-networks.pdf

Database key: helayers-a-tile-tensors-framework-for-large-neural-networks-on · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set helayers-a-tile-tensors-framework-for-large-neural-networks-on COLUMN VALUE