Security Evaluation of Glitch Based Authentication Function for Edge AI

In recent years, the utilization of artificial intelligence (AI) in embedded devices such as edge devices has been expected. NN GPUF, which combines AI (Neural Network: NN) and glitch physically unclonable function (GPUF), an authentication function, has been proposed as a technology to secure edge AI devices. However, modeling attacks against NN GPUFs have not been studied, and its security has not been sufficiently verified. Therefore, this study proposes a modeling attack for NN GPUF. In experiments, a modeling attack using deep neural networks with 100,000 training data set was conducted on a field programmable gate array. As a result, the prediction rate was about 50%, and it was clarified that NN GPUF was resistant to the modeling attack.