Design of Time-Delay Convolutional Neural Networks (TDCNN) Model for Feature Extraction for Side-Channel Attacks

Abstract

This work explores a novel method of SCA profiling to address compatibility problems and bolster Deep Learning (DL) models. Convolutional Neural Networks are proposed in this research as a countermeasure to misalignment-focused countermeasures. We discovered that CNNs provide the potential for end-to-end profiling attacks, where sensitive information can be directly extracted from raw data without the need for any preprocessing. We are of the opinion that dimensionality reduction approaches and realignments both carry the danger of erasing valuable information from data. Indeed, in order to get a well-synchronized dataset, a realignment method modifies signals so that traces are somewhat comparable to one another. "Time-Delay Convolutional Neural Networks" (TDCNN) is more accurate than "Convolutional Neural Network," yet it's still acceptable. It's true that TDCNNs are neural networks based on convolution learned on single spatial information, just as side-channel tracings. However, given to recent surge in popularity of CNNs, particularly from the year 2012 when CNN framework ("AlexNet") achieved Image Net Large Scale Visual Recognition Competition, a notable picture detection competition, moniker TDCNN has been phased out in DL literature. Currently, one needs to employ characteristics related to CNN design, including declaring that one input feature equals 1, for instance, to establish a TDCNN in the most widely used DL libraries.