The Applications Of Data Augmentation Techniques In Deep Learning-based Profiled Side Channel Attacks

As a method to undermine information security,side channel attacks(SCA)exploit physical leakage(such as power consumption,electromagnetic emanation and timing information),acquired during a device running a cryptographic implementation,to recover the key.Among SCA,profiled SCA are the the most powerful ones,which are represented by template attacks(TA).In recent years,deep learning has been applied to profiled SCA,attaining better attack performance compared to classical TA,which means that the key can be recovered with fewer attack traces.Data augmentation(DA),which consists of a set of techniques that increase the size or quality of a training set,has also been emoloyed in SCA,aiming at enhancing attack performance.We find that the applications of DA in profiled SCA still need to be completed.For instance,researchers have paid little attention to augmenting and taking full advantage of the leakage of traces in the frequency domain.In addition,researchers ordinarily assume that profiling traces are sufficient.Actually,the attacker is probably uncapable of collecting adequate profiling traces due to time and resource constraints.Then there arises the issue that the model can merely learn little leakage and perhaps overfits.Therefore,we mainly study two types of applications of DA and propose two schemes to strengthen deep learning-based profiled SCA.1.SpecAugment-based scheme to strengthen deep learning-based profiled SCA.This scheme applies SpecAugment to SCA for the first time,augmenting the leakage information of traces in the time and frequency domain.By short-time Fourier transform,we obtain the representations of traces in the frequency domain.Afterwards,referring to the augmentation policies in SpecAugment,we add frequency masks or time masks to the spectrograms,such that the features learned by the model are robust to partial loss of frequency information or trace information,thus promoting attack performance.2.Mixup-based scheme to strengthen deep learning-based profiled SCA.Considering the restrictions on SCA in the real world,this scheme for the first time employs Mixup to deal with an inadequate profiling set.We follow Mixup to expand the original profiling set in a linear way,in order to narrow the gap between the distributions of the training set and the unknown test set,so that the model can generalize better.Moreover,with the purpose of further adding potential characteristics to the training set and more effectively resisting the prediction perturbation brought by the test set,we propose an improvement to this scheme.That is,to select two traces with different labels to perform Mixup.For both schemes,we conduct correlation power analysis and confirmatory experiments on four public datasets.The results of correlation power analysis indicate that,the generated spectrograms in the first scheme and the generated traces in the second scheme both reveal the leakage,which is consistent with the leakage in the original traces.The results of confirmatory experiments demonstrate that,in the first scheme,compared to the original trace set and the original spectrogram set,the augmented spectrogram set can recover the key with fewer attack traces.In the second scheme,the augmented trace set can effectively enhance attack performance,especially when the original traces are limited.Moreover,the improved second scheme can further bring about performance gain to the attack.