CryptonDL: Encrypted Image Classification Using Deep Learning Models

Adham Helbawy, Mahmoud Bahaa, Alia El Bolock

2023

Abstract

Deep Neural Networks (DNNs) have surpassed traditional machine learning algorithms due to their superior performance in big data analysis in various applications. Fully homomorphic encryption (FHE) contributes to machine learning classification, as it supports homomorphic operations over encrypted data without decryption. In this paper, we propose a deep learning model, CryptonDL, that utilizes TenSEAL’s CKKS scheme to encrypt three image datasets and then classify each encrypted image. This model first trains the image datasets without encryption using a PyTorch convolutional neural network model. Using the weights of the convolutional neural network model in the encrypted convolutional neural network model, each image will be encrypted and then only decrypted in the prediction results. TenSEAL implemented the same model, but this model was optimized to achieve higher accuracy than TenSEAL’s original model. CryptonDL achieved an encrypted image classification accuracy of 98.32 percent and an F1 score of 0.9832 on the MNIST dataset, an accuracy of 88 percent and an F1 score of 0.8811 on the Fashion MNIST, and an accuracy of 92 percent and an F1 score of 0.9207 on the Kuzushiji MNIST. CryptonDL shows that encrypted image classification could be achieved with high accuracy without using pre-trained models.

Paper Citation

in Harvard Style

Helbawy A., Bahaa M. and El Bolock A. (2023). CryptonDL: Encrypted Image Classification Using Deep Learning Models. In Proceedings of the 12th International Conference on Data Science, Technology and Applications - Volume 1: DATA; ISBN 978-989-758-664-4, SciTePress, pages 367-374. DOI: 10.5220/0012088000003541

in Bibtex Style

@conference{data23,
author={Adham Helbawy and Mahmoud Bahaa and Alia El Bolock},
title={CryptonDL: Encrypted Image Classification Using Deep Learning Models},
booktitle={Proceedings of the 12th International Conference on Data Science, Technology and Applications - Volume 1: DATA},
year={2023},
pages={367-374},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0012088000003541},
isbn={978-989-758-664-4},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 12th International Conference on Data Science, Technology and Applications - Volume 1: DATA
TI - CryptonDL: Encrypted Image Classification Using Deep Learning Models
SN - 978-989-758-664-4
AU - Helbawy A.
AU - Bahaa M.
AU - El Bolock A.
PY - 2023
SP - 367
EP - 374
DO - 10.5220/0012088000003541
PB - SciTePress