CINXE.COM
PoPETs Proceedings — HeLayers: A Tile Tensors Framework for Large Neural Networks on Encrypted Data
<?xml version="1.0" encoding="utf-8"?> <!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Strict//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-strict.dtd"> <html xmlns="http://www.w3.org/1999/xhtml"> <head> <title>PoPETs Proceedings — HeLayers: A Tile Tensors Framework for Large Neural Networks on Encrypted Data</title> <style> @import url(../../common/respstyle.css); </style> <meta name="viewport" content="width=device-width,initial-scale=1"/> <meta name="citation_title" content="HeLayers: A Tile Tensors Framework for Large Neural Networks on Encrypted Data" /> <meta name="citation_author" content="Ehud Aharoni" /> <meta name="citation_author" content="Allon Adir" /> <meta name="citation_author" content="Moran Baruch" /> <meta name="citation_author" content="Nir Drucker" /> <meta name="citation_author" content="Gilad Ezov" /> <meta name="citation_author" content="Ariel Farkash" /> <meta name="citation_author" content="Lev Greenberg" /> <meta name="citation_author" content="Ramy Masalha" /> <meta name="citation_author" content="Guy Moshkowich" /> <meta name="citation_author" content="Dov Murik" /> <meta name="citation_author" content="Hayim Shaul" /> <meta name="citation_author" content="Omri Soceanu" /> <meta name="citation_journal_title" content="Proceedings on Privacy Enhancing Technologies" /> <meta name="citation_issn" content="2299-0984" /> <meta name="citation_publication_date" content="2023" /> <meta name="citation_publication_volume" content="2023" /> <meta name="citation_publication_issue" content="1" /> <meta name="citation_publication_firstpage" content="325" /> <meta name="citation_publication_lastpage" content="342" /> <meta name="citation_pdf_url" content="https://petsymposium.org/popets/2023/popets-2023-0020.pdf" /></head> <body> <div class="container"> <img class="banner" src="/images/banner/banner_popets.png"> <div class="nav-main"> <label for="show-menu" class="show-menu">Menu</label> <input type="checkbox" id="show-menu" role="button"> <ul id="menu"> <li><a href="/">Home</a></li> <li><a href="/popets/">Proceedings</a></li> <li><a href=".">Volume 2023</a></li> <li><a href="popets-2023-0019.php">Previous</a></li> <li><a href="popets-2023-0021.php">Next</a></li> </ul> </div> <div class="content"> <h2>HeLayers: A Tile Tensors Framework for Large Neural Networks on Encrypted Data</h2> <p><b>Authors:</b> Ehud Aharoni (IBM Research - Israel), Allon Adir (IBM Research - Israel), Moran Baruch (IBM Research - Israel and Bar Ilan University), Nir Drucker (IBM Research - Israel), Gilad Ezov (IBM Research - Israel), Ariel Farkash (IBM Research - Israel), Lev Greenberg (IBM Research - Israel), Ramy Masalha (IBM Research - Israel), Guy Moshkowich (IBM Research - Israel), Dov Murik (IBM Research - Israel), Hayim Shaul (IBM Research - Israel), Omri Soceanu (IBM Research - Israel)</p> <p><b>Volume:</b> 2023<br /> <b>Issue:</b> 1<br /> <b>Pages:</b> 325–342<br /> <b>DOI:</b> <a href="https://doi.org/10.56553/popets-2023-0020">https://doi.org/10.56553/popets-2023-0020</a></p> <p><a href="https://ibm.github.io/helayers/" target="_blank"><img src="../artifact_badge.png" alt="artifact"></a></p> <p><b><a href="popets-2023-0020.pdf">Download PDF</a></b></p> <p><b>Abstract:</b> Privacy-preserving solutions enable companies to offload confidential data to third-party services while fulfilling their government regulations. To accomplish this, they leverage various cryptographic techniques such as Homomorphic Encryption (HE), which allows performing computation on encrypted data. Most HE schemes work in a SIMD fashion, and the data packing method can dramatically affect the running time and memory costs. Finding a packing method that leads to an optimal performant implementation is a hard task. We present a simple and intuitive framework that abstracts the packing decision for the user. We explain its underlying data structures and optimizer, and propose a novel algorithm for performing 2D convolution operations. We used this framework to implement an inference operation over an encrypted HE-friendly AlexNet neural network with large inputs, which runs in around five minutes, several orders of magnitude faster than other state-of-the-art non-interactive HE solutions.</p> <p><b>Keywords:</b> homomorphic encryption, packing optimization, privacy preserving machine learning, neural networks, non-interactive homomorphic encryption, convolutional layers</p> <p>Copyright in PoPETs articles are held by their authors. This article is published under a <a href="https://creativecommons.org/licenses/by/4.0/">Creative Commons Attribution 4.0</a> license.</p> <a href="popets-2023-0020.pdf"><img src="/2023/files/papers/issue1/popets-2023-0020.png" /></a> </div> <div class="footer"> <div class="nav-footer"> <a href="/contact.php">Contact Us</a> </div> <div class="sns-footer"> <a href="https://youtube.com/channel/UC-m6oi7a-8LffTk64J3tq-w" target="_blank"><img class="sns-icon" src="/images/icons/yt_icon_mono_dark.png"></a> <a href="https://twitter.com/pet_symposium" target="_blank"><img class="sns-icon" src="/images/icons/twitter.png"></a> <a href="https://infosec.exchange/@PET_Symposium" target="_blank"><img class="sns-icon" src="/images/icons/mastodon-logo-white.png"></a> </div> </div> </body> </html>