Approximate Homomorphic Pre-Processing for CNNs

Shabnam Khanna, Ciara Rafferty

2023

Abstract

Homomorphic encryption (HE) allows computations on encrypted data, making it desirable for use in privacy-preserving data analytics. However, HE function evaluation is computationally intensive. Approximate computing (AC) allows a trade-off between accuracy, memory/energy usage and running time. Polynomial approximation of the Rectified Linear Unit (ReLU) function, a key CNN activation function, is explored and AC techniques of task-skipping and depth reduction are applied. The most accurate ReLU approximations are implemented in nGraph-HE’s Cryptonets CNN using a SEAL backend, resulting in a minimal decrease in training accuracy of 0.0011, no change in plaintext classification accuracy, and a speed-up of 47%.

Paper Citation

in Harvard Style

Khanna S. and Rafferty C. (2023). Approximate Homomorphic Pre-Processing for CNNs. In Proceedings of the 20th International Conference on Security and Cryptography - Volume 1: SECRYPT; ISBN 978-989-758-666-8, SciTePress, pages 710-715. DOI: 10.5220/0012085400003555

in Bibtex Style

@conference{secrypt23,
author={Shabnam Khanna and Ciara Rafferty},
title={Approximate Homomorphic Pre-Processing for CNNs},
booktitle={Proceedings of the 20th International Conference on Security and Cryptography - Volume 1: SECRYPT},
year={2023},
pages={710-715},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0012085400003555},
isbn={978-989-758-666-8},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 20th International Conference on Security and Cryptography - Volume 1: SECRYPT
TI - Approximate Homomorphic Pre-Processing for CNNs
SN - 978-989-758-666-8
AU - Khanna S.
AU - Rafferty C.
PY - 2023
SP - 710
EP - 715
DO - 10.5220/0012085400003555
PB - SciTePress