Deep Learning Side Channel Analysis Methods Based On Data Aggregation And Data Augmentation

Side channel analysis is a powerful and highly threatening cryptanalysis technique by capturing and analyzing information leaked during device operation to recover keys.As cryptographic devices become more sophisticated and various safeguards are applied,the efficiency and success rate of traditional side-channel analysis is greatly reduced.To improve the efficiency and success rate of analysis,deep learning techniques are used in side channel analysis.Deep learning-based side channel analysis also faces some problems and challenges,such as complex model training and heavy workload.It requires a large amount of data for training but data collection may be difficult in practice.To address the above problems,this paper conducts side channel analysis based on neural networks,uses multi-source data aggregation and intra-class Cut Mix data enhancement methods to train neural network models,and carries out experimental tests on key recovery of AES algorithm,mainly as follows.1.Deep learning-based side channel analysis require modelling and training for each key byte of the cryptographic algorithm,with high price of data collecting and model training.To address this problem,a side-channel analysis method based on multi-source data aggregation and neural network is proposed.Take the AES＿128 algorithm as an example,in order to screen the leaked data of key byte with good generalization effect for data aggregation,16 single-key byte models are trained based on the leaked data of 16 key bytes to recovery the 16 key bytes respectively at first.Secondly,a scoring mechanism is designed to evaluate the generalization effect of each single-key byte model,and the single-key byte model with the best recovery effect for each key byte is selected by the score ranking.Finally,a multi-source data aggregation model is constructed with the leaked data set of each key byte corresponding to the selected model for training to recovery the key.The experiment results show that the multi-source data aggregation model has good generalization effect,effectively improves the accuracy and efficiency of key recovery,reduces the number of power consumption traces required to recover keys.The model also has a good attack effect with less traces.2.In reality,due to the constraints of collection time,lifecycle of the key and device protection strategy,it is impossible to collect enough power consumption traces to train the model,which leads to poor model performance and affects key recovery.To address this problem,a data augmentation method based on the intra-class Cut Mix method is proposed for side channel analysis.Troughing the same class of constraints,the intra-class Cut Mix solves the problem of corrupting the correlation between power traces and intermediate values in the Cut Mix method.This method firstly augments the training set,and then constructs the MLP neural network for training.It recoveries key for the AES algorithm finally.The experimental results show that the intra-class Cut Mix method generates power traces and performs deep learning side channel analysis,which can help the model learn relevant features and reduce the number of traces required to recover keys.