Performance Evaluation of AI Authentication Device Implemented on SAKURA-G

Abstract

In recent years, AI security issues have been reported. To ensure the authenticity of AI devices, the neural network physically unclonable function (NN PUF) has been proposed. To generate the unique ID for authentication, the NN PUF extracts the variance of calculation time in AI inference. The previous study evaluates the NN PUF by 65nm field programmable gate array (FPGA); however, no evaluation by other process has been performed. As the PUF extracts small variation in LSI, PUF performance may change in different processes, so the evaluation in various processes is important. Therefore, the present study evaluates the NN PUF implemented into 45nm FPGA on SAKURA-G.