Non-Profiled Deep Learning-Based Side-Channel Analysis With Only One Network Training

We propose efficient protocols for non-profiled deep learning-based side-channel analysis (DL-SCA). While the existing protocol, proposed by Timon in 2019, requires computational resources for training as many neural networks as the number of key candidates, our protocol requires training only one network, which can be transformed into a network associated with each key candidate. For instance, in the case of analysis for the AES, the network training complexity is 1/256 of that for the existing protocol. In this study, we describe our idea and formulate it as two protocols depending on the metrics used. We numerically examine them by implementing each protocol with two network architectures, multilayer perceptron and convolutional neural network. Using publicly available open data (ASCAD), we show that both protocols efficiently work as expected. We also clarify that our trained network, as in Timon’s original case, can be recycled for an attack against the same device with different key materials. Non-profiled DL-SCAs are superior to profiled ones in that they require no reference device for profiling before analyzing the target device. This property holds for our proposal as well.