Privacy Preserving Deep Learning Through Fully Homomorphic Encryption and Advanced GPU Acceleration Techniques
The aim of this project is to develop advanced packing and encoding techniques for Fully Homomorphic Encryption to enable parallel execution. The proposed techniques aim to improve efficiency of homomorphic operations and control the growth of noise after every multiplication. It also aims to handle approximation of activation functions, so that the homomorphic operations can be performed without losing too much accuracy. The developed techniques target GPU platform with massively parallel architecture. Techniques for accelerating the homomorphic operations utilizing state-of-the-art GPU like tensor cores and graph-based task scheduling will be proposed in this project. The main focus will be on improving the speed of large integer multiplication, which is the main bottleneck in homomorphic operations.
In this project, we explore the algebraic properties of a p-clean ring which is typically studied in modern algebra theory that can be used in constructing public key encryption schemes. The construction will then be extended to others cryptographic primitives such as signature scheme, key-exchange protocol and zero-knowledge proof. As various algebraic structures serve as a fundamental tool in many cryptography areas, we hope that the outcome of this project can enlighten the concept of p-clean ring into attentions. More precisely, we show that there is a remarkable feedback from cryptography to modern algebra theory because some of the problems motivated by cryptography appear to be new to p-clean ring, and they open many interesting research avenues within ring and group ring theory. On the other hand, the security analysis of the proposed scheme will fully utilize the internal structures of the p-clean ring together with systematic complexity argument, we aim to prove that our proposed scheme is secure under indistinguishable adaptively chosen ciphertext attack (IND-CCA2). Based on our understanding, we also strongly believe that the study of this direction of cryptography will instil many open problems and new research directions to suit current advanced usage of cryptography and security issues in IoT devices, cloud storage, blockchain, IR4.0 and even mechanism in data science related applications.