Intel, Microsoft Push Homomorphic Encryption with Open-Source Moves

Earlier this month, Intel open-sourced HE-Transformer, the tool that frees artificial intelligence models to run on encrypted data. The move pioneers new territory for learning and analysis, while operating alongside complementary actions taken by industry peers.

Sensitive information, including but not limited to that generated by the healthcare industry, can be used as a dataset to feed AI models, allowing valuable insights to be gleaned without compromising the underlying personal information.

“HE allows computation on encrypted data” Intel research scientist Opens a new window Fabian Boemer Opens a new window and Casimir Wierzynski, senior research director,Opens a new window explained in a blog post. “This capability, when applied to machine learning, allows data owners to gain valuable insights without exposing the underlying data; alternatively, it can enable model owners to protect their models by deploying them in encrypted form.”

The chipmaker made the announcementOpens a new window at a conference on neural information processing systems in Montreal, Canada, a few days after Microsoft Research said it was open-sourcingOpens a new window its Simple Encrypted Arithmetic Library (SEAL) encryption library on software development platform GitHub under an MIT license.

HE-Transformer is a backend for Intel’s nGraph neural network compiler and uses SEAL to implement the underlying cryptography functions.

“We are excited to work with Intel to help bring homomorphic encryption to a wider audience of data scientists and developers of privacy-protecting machine learning systems” said Kristin Lauter, principal researcher and research manager of Microsoft Research’s cryptography group.

HE Standardization Push

HE stands for homomorphic encryption, a type of cryptography that allows computational analysis of plaintext files that have been encrypted, also known as ciphertexts. Computations can be carried out without the need to decrypt the data. The results are encrypted and can only be seen in full by the owner of the decryption key.

First developed in 2009 by IBM researcher Craig Gentry, early HE processes required excessive time and processing power. IBM followed the development with the launch of its own HE library named Helib in 2013. In the years since the launch, processing efficiency has been greatly improved.

Microsoft’s SEAL was unveiled in 2015 and was written in standard C++ programming language, working on Windows, Linux and OS X. It incorporates both the Brakerski/Fan-Vercauteren (BFV) and the Cheon-Kim-Kim-Song (CKKS) encryption schemes.

A prominent group of industry leaders have been pushing to have HE standardized, with Microsoft, Intel, IBM and SAP acting as key members of the Homomorphic Encryption Standardization group. Microsoft’s release of SEAL to open source comes amid this broader industry effort to push forward HE.

End to Encryption Key Vulnerability

“With traditional encryption schemes, it is impossible to run any computation on encrypted data,” write Microsoft researchers Lauter, Kim Laine and Sreekanth Kannepalli. “So either we store our data encrypted in the cloud and download it to perform any useful operations, which can be logistically inconvenient, or we provide the decryption key to service providers, risking our privacy. Until now. Homomorphic encryption, which allows processing of encrypted data, gives us the ability to use these services without exposing our private information.”

As the Intel researchers point out, the design of deep-learning HE models requires expertise in deep learning, encryption and in software engineering. Intel’s HE-Transformer tool allows data scientists to deploy trained models using machine learning frameworks such as Google’s TensorFlowOpens a new window , with plans to integrate with Facebook’s PyTorch as well.

Users of deep-learning frameworks including PyTorch, CNTK and MXNet that can export neural networks to open source AI ecosystem Open Neural Network Exchange, can import their model to HE-Transformer using the ONNX importer in nGraph, then export it in a serialized format, ready to run.

The growing prevalence of internet-enabled devices and the fast-rising amounts of data they generate accompany a need for more data security.

Vested interest or not, Intel maintains that increasing privacy concerns makes HE an attractive solution in the struggle to balance the need for datasets to feed machine learning, and the privacy requirements that would prevent their use.