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

Homomorphic encryption (HE) ---the ability to perform computation on encrypted---is an attractive remedy to increasing concerns about data privacy in learning (DL). However, building DL models that operate on ciphertext is labor-intensive and requires simultaneous expertise in DL, , and software engineering. DL frameworks and recent advances in compilers have greatly accelerated the training and deployment of DL to various computing platforms. We introduce nGraph-HE, an extension of, Intel's DL graph compiler, which enables deployment of trained models popular frameworks such as TensorFlow while simply treating HE as another target. Our graph-compiler approach enables HE-aware optimizations-- at compile-time, such as constant folding and HE-SIMD packing, and run-time, such as special value plaintext bypass. Furthermore, nGraph-HE with DL frameworks such as TensorFlow, enabling data scientists to DL models with minimal overhead.