As part of this commitment, we open-sourced a first-of-its-kind Fully Homomorphic Encryption (FHE) transpiler two years ago, and have continued to remove barriers to entry along the way. FHE is a powerful technology that allows you to perform computations on encrypted data without being able to access sensitive or personal information and we’re excited to share our latest developments that were born out of collaboration with our developer and research community to help expand what can be done with FHE.
Today, we are introducing new tools that enable anyone to apply FHE technologies to video files. This advancement is important because video adoption can often be expensive and incur long run times, limiting the ability to scale FHE use to larger files and new formats.
To do so, we are expanding our FHE toolkit in three new ways to make it easier for developers to use FHE for a wider range of applications, such as private machine learning, text analysis, and the aforementioned video processing. As part of our toolkit, we are releasing new hardware, a software crypto library and an open source compiler toolchain. Our goal is to provide these new tools to researchers and developers to help advance how FHE is used to protect privacy while simultaneously lowering costs.
Link to Youtube Video (visible only when JS is disabled)
We believe—with more optimization and specialty hardware — there will be a wider amount of use cases for a myriad of similar private machine learning tasks, like privately analyzing more complex files, such as long videos, or processing text documents. Which is why we are releasing a TensorFlow-to-FHE compiler that will allow any developer to compile their trained TensorFlow Machine Learning models into a FHE version of those models.
Once a model has been compiled to FHE, developers can use it to run inference on encrypted user data without having access to the content of the user inputs or the inference results. For instance, our toolchain can be used to compile a TensorFlow Lite model to FHE, producing a private inference in 16 seconds for a 3-layer neural network. This is just one way we are helping researchers analyze large datasets without revealing personal information.
In addition, we are releasing Jaxite, a software library for cryptography that allows developers to run FHE on a variety of hardware accelerators. Jaxite is built on top of JAX, a high-performance cross-platform machine learning library, which allows Jaxite to run FHE programs on graphics processing units (GPUs) and Tensor Processing Units (TPUs). Google originally developed JAX for accelerating neural network computations, and we have discovered that it can also be used to speed up FHE computations.
Finally, we are announcing Homomorphic Encryption Intermediate Representation (HEIR), an open-source compiler toolchain for homomorphic encryption. HEIR is designed to enable interoperability of FHE programs across FHE schemes, compilers, and hardware accelerators. Built on top of MLIR, HEIR aims to lower the barriers to privacy engineering and research. We will be working on HEIR with a variety of industry and academic partners, and we hope it will be a hub for researchers and engineers to try new optimizations, compare benchmarks, and avoid rebuilding boilerplate. We encourage anyone interested in FHE compiler development to come to our regular meetings, which can be found on the HEIR website.
Organizations and governments around the world continue to explore how to use PETs to tackle societal challenges and help developers and researchers securely process and protect user data and privacy. At Google, we’re continuing to improve and apply these novel techniques across many of our products, through our Protected Computing, which is a growing toolkit of technologies that transforms how, when and where data is processed to technically ensure its privacy and safety. We’ll also continue to democratize access to the PETs we’ve developed as we believe that every internet user deserves world-class privacy.
