Deep CNN Implemented on Fully Homomophic Encrypted Data

Overview

A deep convolutional neural network (CNN) has been implement with Simple Encrypted Arithmetic Library - SEAL. A classical image classfication example has been used to demonstrates the capability of CNN-based inference on fully homomophic encrpted (FHE) handwritted digts from the MNIST database. The regular CNN model based on Keras has been developed to generate the weights. The Python implementation is also included.

Requirements

• Simple Encrypted Arithmetic Library - SEAL
• Visual Studio Community 2017
• Python 2.6 or up or Python 3

Installation

• Install SEAL
• INSIDE the SEAL folder, clone this respo. This step will create a project folder named "FHE_CNN" insdie SEAL
• Open the SEAL project with VS 2017, and add the folder "FHE_CNN" as an exsiting project
• Set the project "FHE_CNN" as the startup project