Neural Networks and Security

Research Overview

Research is advancing in fully homomorphic encryption (FHE), a technology that can perform all operations (both arithmetic and logical) in the encrypted domain, as an extension of advanced cryptography. Although currently it requires extensive execution time even with high-performance supercomputers, making it impractical, it is a promising technology for the future. Implementation on high-performance hardware is expected to accelerate the practical application of FHE. At Ikeda Lab, we are advancing research for the optimal hardware implementation focusing on the versatile implementation of TFHE, hardware implementation limited to additive homomorphic properties, and for specific purposes, the implementation of homomorphic neurons in neural networks that operate while protecting data, parameters, or both.

Acceleration of TFHE

Applications of cloud computing that protect data privacy using fully homomorphic encryption (HE) are expected. However, the computational cost of HE is 1000 to 10000 times higher than plaintext, and practical applications have not yet been realized. We are working on accelerating the computation of HE from both software and hardware aspects. In software, we modify traditional neural networks to structures suitable for HE and reduce computation costs through preprocessing. Additionally, we design dedicated hardware to accelerate computationally intensive operations on the software side. By speeding up the process through both software and hardware, we aim to develop practical applications using HE.