Skip to content

TenSEAL: A Library for Encrypted Tensor Operations Using Homomorphic Encryption

Authors: Ayoub Benaissa; Bilal Retiat; Bogdan Cebere; Alaa Eddine Belfedhal
Year: 2021
Venue: arXiv 2021

type: library collection: catalogue status: auto

Links: arXiv:2104.03152

Summary

Auto-extracted summary (unreviewed)

TenSEAL is an open-source C++/Python library (OpenMined) that exposes encrypted tensor operations on top of the CKKS implementation in Microsoft SEAL, converting tensors from PyTorch or TensorFlow into encrypted vectors and handling context, keys and client-server serialization via Protocol Buffers. Evaluating a small CNN (one 7x7 convolution, two linear layers, square activations) on encrypted MNIST images takes 887 ms on a 16-vCPU AWS instance with 427 KB of communication and 97.4% accuracy versus 97.7% in the clear.

Abstract

Machine learning algorithms have achieved remarkable results and are widely applied in a variety of domains. These algorithms often rely on sensitive and private data such as medical and financial records. Therefore, it is vital to draw further attention regarding privacy threats and corresponding defensive techniques applied to machine learning models. In this paper, we present TenSEAL, an open-source library for Privacy-Preserving Machine Learning using Homomorphic Encryption that can be easily integrated within popular machine learning frameworks. We benchmark our implementation using MNIST and show that an encrypted convolutional neural network can be evaluated in less than a second, using less than half a megabyte of communication.
Source: arxiv

Extracted details

Field Value Source Confidence
contribution open-source library for encrypted tensor computation using CKKS in Microsoft SEAL llm 0.90
contribution direct conversion of PyTorch / TensorFlow tensors to encrypted tensors llm 0.90
contribution context, plain-tensor and encrypted-tensor API with Protocol Buffers client-server exchange llm 0.90
contribution im2col convolution and dot-product algorithms for encrypted CNN evaluation llm 0.90
contribution sub-second encrypted MNIST CNN inference with under half a megabyte of communication llm 0.90
schemes CKKS llm 0.90
libraries TenSEAL llm 0.90
libraries SEAL llm 0.90
libraries PyTorch llm 0.90
libraries TensorFlow llm 0.90
models CNN (1 conv layer + 2 linear layers, square activation) llm 0.90
datasets MNIST llm 0.90
tasks encrypted tensor operations llm 0.90
tasks private CNN inference llm 0.90
tasks image classification llm 0.90
hardware AWS c4.2xlarge (8 vCPUs, 15 GiB) llm 0.90
hardware AWS c4.4xlarge (16 vCPUs, 30 GiB) llm 0.90
security level 128-bit llm 0.90
threat model client-server MLaaS; server evaluates on encrypted input and never sees input or output llm 0.90
interaction non-interactive FHE llm 0.90
code url https://github.com/Microsoft/SEAL (URL found in the full text) regex 0.50
code url https://github.com/protocolbuffers/protobuf (URL found in the full text) regex 0.50
code url https://github.com/OpenMined/TenSEAL (URL found in the full text) regex 0.50
code url https://github.com/OpenMined/TenSEAL llm 0.90
llm relevance medium llm 0.90
key result full forward latency: 887.06 ms (encrypted MNIST CNN, AWS c4.4xlarge 16 vCPUs) llm 0.90
key result full forward latency: 1456.29 ms (encrypted MNIST CNN, AWS c4.2xlarge 8 vCPUs) llm 0.90
key result accuracy: 97.4% encrypted vs 97.7% plaintext (MNIST test set) llm 0.90
key result communication: 427 KB (send encrypted input and receive encrypted output) llm 0.90
key result parameters: N=8192, 206-bit coefficient modulus, 21-bit scale (CKKS, 6 multiplications, 128-bit security) llm 0.90

System card: TenSEAL

Group 7. FHE libraries, GPU back-ends and compilers
Family CKKS, BFV (SEAL)
Arithmetic Python
Interaction ML-on-HE toolkit
Model scale -
Status Niche
Relation to the SoK Easy Python tensor API over SEAL.
Note Good for teaching and prototypes. No bootstrapping.
Code https://github.com/OpenMined/TenSEAL
Systems page TenSEAL

Keyword mentions

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

  • scheme: CKKS 12 [p.2], Leveled / somewhat HE 2 [p.3]
  • library: TenSEAL 12 [p.1], Microsoft SEAL 4 [p.2], PyTorch 2 [p.2]
  • model: CNN 4 [p.2], MobileNet 1 [p.4]
  • dataset: MNIST 6 [p.1]
  • hardware: GPU 2 [p.4], Intel Xeon 1 [p.12]
  • technique: Convolution 22 [p.2], Matrix multiplication 7 [p.3], Rotations 5 [p.2], Relinearization 3 [p.2]
  • topic: Homomorphic encryption 16 [p.1], Benchmark 6 [p.1], Open source 5 [p.1], Privacy-preserving machine learning 3 [p.1]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/catalogue/7-109_tenseal-a-library-for-encrypted-tensor-operations-using-homo.pdf catalogue 12 375.7 KB 45ed2a786abb

Database key: tenseal-a-library-for-encrypted-tensor-operations-using-homomorphic · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set tenseal-a-library-for-encrypted-tensor-operations-using-homomorphic COLUMN VALUE