Security Attack on Remote Sensing Equipment: PoIs Recognition Based on HW with Bi-LSTM Attention

Abstract

Deep learning is an influencer in hardware security applications, which grows up to be an essential tool in hardware security, threats the confidentiality, integrity, and availability of remote sensing equipment. Comparing to traditional physical attack, not only it can greatly reduce the workload of manual selection of POIs (Points of Interests) in security attack and Trojan backdoor, but also replenishes the toolbox for attacking. On account of minute changes between network structure model and hyperparameters constantly affecting the training and attacking effect, literally, deep learning serves as a tool but not key role in hardware security attack, which means it cannot completely replace template attack and other traditional energy attack methods. In this study, we present a method using Bi-LSTM Attention mechanism to focus on the POIs related to Hamming Weight at the last round s-box output. Firstly, it can increase attacking effect and decrease guessing entropy, where attacking FPGA data demonstrates the efficiency of attacking. Secondly, it is different from the traditional template attack and deep learning attack without preprocessing subjecting to raw traces but provides attentional POIs which is the same with artificial selection. Finally, it provides a solution for attacking encrypting equipment running in parallel.