SHF: Small: Application-Aware Homomorphic Encryption for Neural Networks: From Algorithm to Hardware Accelerators

As the amount of data being generated is rapidly increasing and many technologies keep scaling up, more and more applications are relying on cloud servers for high-performance computing and data processing. The scalability offered by cloud computing allows users to meet performance goals without maintaining expensive infrastructures. In particular, the booming applications of artificial intelligence and increasingly complicated neural networks for higher precision make offering inference services and/or training neural networks a major application of cloud computing.

In many cases, the user data contains private information, such as patient images for medical diagnosis, genome samples for sequencing, and financial data for analysis. User data can be protected against eavesdropping by cryptography schemes during the transmission to and from cloud servers. However, if traditional cryptography schemes are utilized, decryption must be carried out before any computation is possible and the cloud server will get access to user private data.

Privacy-preserving cloud computing and machine learning are enabled by the new homomorphic encryption (HE) technique, which allows computations to be directly carried out on encrypted data. Using HE, the cloud server does not gain any information on the user data and the computation results are encrypted. However, the achievable speed of HE implementations is far from being practical despite previous research efforts.

This project aims to speed up HE implementations by orders of magnitude, expanding the privacy-preserving capabilities of cloud providers.

This project pursues scalability improvements by taking into account the specific computations involved in applications and integrating algorithmic reformulations with hardware architecture design. Such application-aware cross-layer design approaches can enable unprecedented complexity reduction. For the first time, new HE operators implementing combined computations with much reduced complexity will be investigated.

In addition, the overall complexity of homomorphically encrypted neural networks will be further reduced by developing new techniques that enable the sharing of intermediate results and more efficient packing of data into ciphertexts. The algorithmic reformulation, new operator design, and computation optimization are carried out jointly with the corresponding hardware architecture design to truly speed up HE in real implementations.

As a result of this project, HE hardware implementations achieving practical speed with low complexity will be developed for deep neural networks. The new designs can be also extended to other domains and will have significant impacts on privacy-preserving medical diagnoses, genome sequencing, data analytics, and many other applications.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.