CRII: SHF: RUI: Custom Hardware Accelerators for Privacy-Preserving Image Processing

Homomorphic encryption is an important cryptographic technique that allows direct computation on encrypted data without decryption. The result of the computation is entirely encrypted, and only the data owner can decrypt it using a private key. As a result, data analysts such as cloud-service providers cannot view any private or sensitive information from the data.

Promising applications of homomorphic encryption include financial and medical data analytics and privacy-preserving machine learning, and homomorphic encryption can be also used for more diverse applications such as genomics, national security/critical infrastructures, and elections. Several emerging applications makes homomorphic encryption increasingly essential.

However, homomorphic encryption suffers from slow processing speed, which renders it impractical for many critical applications. In addition, it only supports addition and/or multiplication for encrypted data, which limits the scope of its applications. This project will address these issues using custom hardware accelerators and a numerical approach that approximates several arithmetic and logical operations using addition and multiplication.

This project will broaden the scope of practical homomorphic-encryption-based applications and provide opportunities for students to gain experience in a variety of areas such as digital system design, software programming, and cryptography.

This project will focus on homomorphic-encryption-based image-processing applications. In typical homomorphic encryption-based systems, a user encrypts private data on a local device and sends them to a cloud server where a homomorphic encryption-based application is performed. However, the inclusion of noise or brightness distortion in the original images can negatively affect the results of the application.

Therefore, one of the goals of this project is to develop custom hardware accelerators for homomorphic-encryption-based noise cancelling and contrast enhancement using field-programmable gate arrays. They require division and min/max functions for encrypted data, and these operations are being numerically implemented. Another problem arises when data analysts convey homomorphic-encryption-driven analysis results to a user.

They can only refer to specific data points associated with a section of the image, but not the original image itself as it is encrypted. This can be problematic on the user's side as they may not understand which part of the image the data analyst is referring to. Thus, another goal is to develop a custom hardware accelerator for an image-thresholding technique for highlighting regions of interest.

It requires comparison operations for encrypted data, and the numerical approach is being used again. This project aims at reducing the computational runtime overhead by one order to two orders of magnitude compared to software implementations.

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.