Research On Methods Of Power Traces Pre-processing Based On Convolutional Neural Networks

Recently,with the rapid development of intelligence information technology,peo-ple are further inter connected with various intelligent devices and embedded devices.These devices are required to comprise security models,the fundamental of which is cryptography.Side-channel attack is a practical cryptoanalysis comparing to classical algebraic analysis,and faces the security of the implementation of cryptographic tech-nology.This kind of attack brings great threats to secure systems and thus becomes an attractive issue of more and more cryptoanalysis researchers.In a common side-channel attack,attackers collect the timing behavior,power con-sumption,electromagnetic emanation,acoustic emissions or thermal effects of the tar-get device running a cryptographic algorithm with an employed acquisition equipment.Then,they analyze the relationship between the collected data and intermediate values during the device's execution.Since the intermediate values usually depend on the key of the cryptographic algorithm,attackers could recover the key based on the analysis results.The collected data are usually named as power traces.As the quality of power traces greatly affects the success rate,pre-processing is required before the analysis performing.A major factor that affects the quality of power traces is key-independent noise.Hence,we need to study effective methods to filter the noise of power traces and improve the relevance between power traces and the key to reduce the complexity of side-channel attacks.Previous work shows that the convolutional neural networks have good performance in handling end-to-end side-channel attacks while it lacks applica-tions in pre-processing power traces.Thus,the thesis researches on such applications and proposes two schemes that pre-process power traces and reduce the noise based on convolutional neural networks.One is a power traces pre-processing scheme based on SincNet convolutional neu-ral network.We implement the Sinc convolution layer called SincConv1D based on the sinc band-pass filter to construct a SincNet network for side channel attacks.Each sinc filter of the Sinc convolutional layer can adaptively learn the high and low cut-off frequencies of power traces,and filter out the noisy bands to provide higher key related features for subsequent network layers.We ultimately achieve the goal of performing the side-channel attack successfully with fewer power traces.The other is a power traces pre-processing scheme based on Sinc convolutional denoising auto-encoder.In the thesis,the Sinc layer and traditional convolutional layers are combined to construct a convolutional denoising auto-encoder.It aims to train the power traces pre-processing model from noisy power traces to clean power traces.In addition,the scheme can effectively lower the influence of noise in power traces and reduce the complexity of side-channel attacks.Finally,experiments are carried out on the public side-channel dataset for the above two schemes.The experimental results show the effectiveness of the proposed schemes.