Welcome to FHEON Documentation
FHEON is an open-source framework for developing privacy-preserving convolutional neural networks using homomorphic encryption. Built on CKKS and OpenFHE, FHEON enables encrypted neural network inference and training while keeping sensitive data private throughout computation.
Designed for both researchers and developers, FHEON provides a simple, modular, and extensible framework for building, evaluating, and deploying HE-friendly deep learning models without requiring expertise in low-level cryptography and machine learning.
Highlights
Privacy-Preserving Inference for encrypted neural network evaluation, on single inputs and batches.
Encrypted Training and Fine-Tuning with encrypted parameters, in the standard layout.
Configurable CNN Layers with familiar deep learning parameters.
Easy Extensibility through a modular architecture.
Easy Model Design using a clean and intuitive API.
Standard and Diagonal Controllers: standard and channel-diagonal (FHEON-D) primitives behind one interface for single inputs, and a multithreaded standard controller for batches.
FHEON Studio, a web workbench that builds and runs the FHEON programs on encrypted MNIST, CIFAR-10 and CIFAR-100 images (FHEON Studio).
Examples for MLP, LeNet-5, VGG-11, VGG-16, ResNet-20, and ResNet-34 on MNIST, CIFAR-10, and CIFAR-100, including multiple optimized model variants and ResNet-20 fine-tuning.
FHEON makes homomorphic encryption more accessible by providing a practical and flexible platform for developing secure deep learning applications.
Citation
If you use FHEON in your work, please cite the following papers:
Encrypted Single-input Inference
@misc{njungle2025fheonconfigurableframeworkdeveloping,
title={FHEON: A Configurable Framework for Developing Privacy-Preserving Neural Networks Using Homomorphic Encryption},
author={Nges Brian Njungle and Eric Jahns and Michel A. Kinsy},
year={2025},
eprint={2510.03996},
archivePrefix={arXiv},
primaryClass={cs.CR},
url={https://arxiv.org/abs/2510.03996}
}
Also available online at FHEON: A Configurable Framework for Developing Encrypted Privacy-Preserving Neural Networks Using Homomorphic Encryption
Encrypted High-throughput Inference
@misc{njungle2026deepencryptedtraininglowlatency,
title={Towards Deep Encrypted Training: Low-Latency, Memory-Efficient, and High-Throughput Inference for Privacy-Preserving Neural Networks},
author={Nges Brian Njungle and Eric Jahns and Michel A. Kinsy},
year={2026},
eprint={2604.16834},
archivePrefix={arXiv},
primaryClass={cs.CR},
url={https://arxiv.org/abs/2604.16834}
}
Also available online at Towards Deep Encrypted Training: Low-Latency, Memory-Efficient, and High-Throughput Inference for Privacy-Preserving Neural Networks
CONTENT
- Introduction
- Getting Started with FHEON
- Binary Usage and Configuration
- FHEON Studio
- Security Levels, Ring Dimension and Packing
- FHEON’s Controllers
- Standard Primitives
- Diagonal Primitives (FHEON-D)
- Encrypted Training
- Evaluation
- Support Functions
- Single Input Examples
- Basic MLP FHE Implementation
- LeNet-5 FHE Implementation
- ResNet-20 FHE Implementation
- ResNet-20 FHE Implementation With Data Preprocessing
- ResNet-20 With Multi-Channel Striding Processing
- ResNet-20 With Multi-Channel Striding and Slot Management
- ResNet-20 Optimized Implementation
- ResNet-34 FHE Implementation
- ResNet-34 Optimized Implementation
- VGG-11 FHE Implementation
- VGG-11 Optimized Implementation
- VGG-16 FHE Implementation
- VGG-16 Optimized Implementation
- Diagonal Examples
- Batch Input Examples
- Training and Fine-Tuning Examples
- Published Papers and Related Projects
- Contact and Community