Research On Neural Network Optimization Method For Side Channel Profiled Attack

Side-Channel Attack is a technology of cryptographic attack that exploits the physical leakages of the encryption process of the cryptographic device to recover the secret key.The physical leakages are the timing,power consumption,electromagnetic(EM)emanation,and even sound.The profiled attack is the most powerful attack of the Side-Channel Attack.It assumes that the attacker has a copy of the target device,but the secret keys are different.Therefore,the attacker can build the attacking model before the actual attack.The attacking model which studies most is the neural network,especially convolutional neural networks(CNN).CNNs have some advantages over the traditional attacking model.CNNs can break through the implementation of encryption algorithms with countermeasures like the mask or random delay.Besides,CNNs can ignore the preprocess as selecting points of interest.Therefore,The performance of CNNs is stronger than the traditional attacking model in most cases.Previous researchers believe that the convolutional neural network has some ability to adapt to noise,but we found that the performance of CNNs is significantly worse in the power trace that has large electronic noise.It indicates that the network model is sensitive to the electronic noise of the energy trace.Therefore,in this work,we optimize the CNN and strengthen the features.The proposed CNNs can reduce the effectiveness of electronic noise without preprocessing like denoising.We explore a neural network optimization scheme for profiled neural network attacks,and then we propose three methods.The optimization of the profiled neural network based on attention mechanism.This method uses the Convolutional Block Attention Mechanism(CBAM).CBAM makes the CNN take more attention to the physical leakages of the power trace and reduces the study to the non-leakage regions.Therefore,it can reduce the effect of the noise in the non-leakage area on CNN,and improve the CNN performance.The optimization of the profiled neural network based on wavelet transform.In the power trace,the data and operation components are generally at low frequencies,while noise components are at high frequencies.So we use discrete wavelet transform to decompose features into high-frequency features and low-frequency and then continually better learn the information related to the operation data.The optimization of the profiled neural network based on DWT(Discrete Wavelet Transformation)attention mechanism.Considering that there are still some limitations to use the attention mechanism and discrete wavelet transform alone,we combine the attention mechanism and discrete wavelet transform to further enhance the expression of features.The experiments show that this method is more effective than the previous in most cases.We verify these three schemes on a dataset with high electronic noise.The experiments show that all of them improve the performance of CNN.Besides,we conducted experiments on three different datasets with/without countermeasures that have less noise,and the ability of the network to attack also has improved.Finally,we visualize these three networks,and the results show that they all learned about the leak area.