Paper 2019/1242 Original: eprint.iacr.org/2019/1242 ↗

Non-Profiled Side Channel Attack based on Deep Learning using Picture Trace

Abstract

In this paper, we suggest a new format for converting side channel traces to fully utilize the deep learning schemes. Due to the fact that many deep learning schemes have been advanced based on MNIST style datasets, we convert from raw-trace based on float or byte data to picture-formatted trace based on position. This is induced that the best performance can be acquired from deep learning schemes. Although the overfitting cannot be avoided in our suggestion, the accuracy for validation outperforms to previous results of side channel analysis based on deep learning. Additionally, we provide a novel criteria for attack success or fail based on statistical confidence level rather than rule of thumb. Even though the data storage is slightly increased, our suggestion can completely be recovered the correct key compared to previous results. Moreover, our suggestion scheme has a lot of potential to improve side channel attack.