Approximate homomorphic pre-processing for CNNs

Shabnam Khanna; Ciara Rafferty

doi:10.5220/0012085400003555

Approximate homomorphic pre-processing for CNNs

Shabnam Khanna, Ciara Rafferty

School of Electronics, Electrical Engineering and Computer Science

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

17 Downloads (Pure)

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 andAC 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 intraining accuracy of 0.0011, no change in plaintext classification accuracy, and a speed-up of 47%.

Original language	English
Title of host publication	Proceedings of the 20th International Conference on Security and Cryptography, SECRYPT 2023
Editors	Sabrina De Capitani di Vimercati, Pierangela Samarati
Publisher	SciTePress
Pages	710-715
ISBN (Electronic)	9789897586668
ISBN (Print)	9781713876496
DOIs	https://doi.org/10.5220/0012085400003555
Publication status	Published - 12 Jul 2023
Event	20th International Conference on Security and Cryptography - Rome, Italy Duration: 10 Jul 2023 → 12 Jul 2023 https://secrypt.scitevents.org/

Publication series

Name	SECRYPT Proceedings
ISSN (Electronic)	2184-7711

Conference

Conference	20th International Conference on Security and Cryptography
Abbreviated title	SECRYPT 2023
Country/Territory	Italy
City	Rome
Period	10/07/2023 → 12/07/2023
Internet address	https://secrypt.scitevents.org/

Keywords

homomorphic evaluation
approximate computing
ReLU
Convolutional Neural Network (CNN)

Access to Document

10.5220/0012085400003555Licence: CC BY-NC-ND

Approximate homomorphic pre-processing for CNNs
Copyright 2023 The Authors. This is an open access article published under a Creative Commons Attribution-NonCommercial-NoDerivatives License (https://creativecommons.org/licenses/by-nc-nd/4.0/), which permits distribution and reproduction for non-commercial purposes, provided the author and source are cited.
Final published version, 2.6 MBLicence: CC BY-NC-ND

20th International Conference on Security and Cryptography
Shabnam Khanna (Participant)
10 Jul 2023 → 12 Jul 2023
Activity: Participating in or organising an event types › Participation in conference

Approximating homomorphic evaluation: an analysis of approximation and optimisation techniques for accelerating homomorphic encryption
Khanna, S. (Author), Rafferty, C. (Supervisor) & O'Neill, M. (Supervisor), Jul 2024
Student thesis: Doctoral Thesis › Doctor of Philosophy

File

Cite this

@inproceedings{bb7bbf61aa194c21a7efa00b1ccedd44,

title = "Approximate homomorphic pre-processing for CNNs",

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 andAC techniques of task-skipping and depth reduction are applied. The most accurate ReLU approximations are implemented in nGraph-HE{\textquoteright}s Cryptonets CNN using a SEAL backend, resulting in a minimal decrease intraining accuracy of 0.0011, no change in plaintext classification accuracy, and a speed-up of 47%.",

keywords = "homomorphic evaluation, approximate computing, ReLU, Convolutional Neural Network (CNN)",

author = "Shabnam Khanna and Ciara Rafferty",

year = "2023",

month = jul,

day = "12",

doi = "10.5220/0012085400003555",

language = "English",

isbn = "9781713876496",

series = "SECRYPT Proceedings",

publisher = "SciTePress",

pages = "710--715",

editor = "{De Capitani di Vimercati}, Sabrina and Pierangela Samarati",

booktitle = "Proceedings of the 20th International Conference on Security and Cryptography, SECRYPT 2023",

note = "20th International Conference on Security and Cryptography, SECRYPT 2023 ; Conference date: 10-07-2023 Through 12-07-2023",

url = "https://secrypt.scitevents.org/",

}

Khanna, S & Rafferty, C 2023, Approximate homomorphic pre-processing for CNNs. in S De Capitani di Vimercati & P Samarati (eds), Proceedings of the 20th International Conference on Security and Cryptography, SECRYPT 2023. SECRYPT Proceedings, SciTePress, pp. 710-715, 20th International Conference on Security and Cryptography, Rome, Italy, 10/07/2023. https://doi.org/10.5220/0012085400003555

Approximate homomorphic pre-processing for CNNs. / Khanna, Shabnam; Rafferty, Ciara.
Proceedings of the 20th International Conference on Security and Cryptography, SECRYPT 2023. ed. / Sabrina De Capitani di Vimercati; Pierangela Samarati. SciTePress, 2023. p. 710-715 (SECRYPT Proceedings).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

TY - GEN

T1 - Approximate homomorphic pre-processing for CNNs

AU - Khanna, Shabnam

AU - Rafferty, Ciara

PY - 2023/7/12

Y1 - 2023/7/12

N2 - 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 andAC 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 intraining accuracy of 0.0011, no change in plaintext classification accuracy, and a speed-up of 47%.

AB - 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 andAC 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 intraining accuracy of 0.0011, no change in plaintext classification accuracy, and a speed-up of 47%.

KW - homomorphic evaluation

KW - approximate computing

KW - ReLU

KW - Convolutional Neural Network (CNN)

U2 - 10.5220/0012085400003555

DO - 10.5220/0012085400003555

M3 - Conference contribution

SN - 9781713876496

T3 - SECRYPT Proceedings

SP - 710

EP - 715

BT - Proceedings of the 20th International Conference on Security and Cryptography, SECRYPT 2023

A2 - De Capitani di Vimercati, Sabrina

A2 - Samarati, Pierangela

PB - SciTePress

T2 - 20th International Conference on Security and Cryptography

Y2 - 10 July 2023 through 12 July 2023

ER -

Approximate homomorphic pre-processing for CNNs

Abstract

Publication series

Conference

Keywords

Access to Document

Fingerprint

Activities

20th International Conference on Security and Cryptography

Student theses

Approximating homomorphic evaluation: an analysis of approximation and optimisation techniques for accelerating homomorphic encryption

Cite this