nGraph-HE: A Graph Compiler for Deep Learning on Homomorphically Encrypted Data

Homomorphic encryption (HE)--the ability to perform computations on encrypted data--is an attractive remedy to increasing concerns about data privacy in the field of machine learning. However, building models that operate on ciphertext is currently labor-intensive and requires simultaneous expertise in deep learning, cryptography, and software engineering. Deep learning frameworks, together with recent advances in graph compilers, have greatly accelerated the training and deployment of deep learning models to a variety of computing platforms. Here, we introduce nGraph-HE, an extension of the nGraph deep learning compiler, which allows data scientists to deploy trained models with popular frameworks like TensorFlow, MXNet and PyTorch directly, while simply treating HE as another hardware target. This combination of frameworks and graph compilers greatly simplifies the development of privacy-preserving machine learning systems, provides a clean abstraction barrier between deep learning and HE, allows HE libraries to exploit HE-specific graph optimizations, and comes at a low cost in runtime overhead versus native HE operations.