| Security technology weakens the strong correlation between side channel information and private data through hiding,masking and combined protection,and is the main method to ensure the security of cryptographic algorithms and circuits.Now,domestic and foreign researchers apply typical attacking techniques to verify the security of protected cryptographic algorithms,to improve the design level of security techniques,to achieve the purpose of "promoting defense with attacking".Therefore,this thesis conducts research on the key technologies of side-channel attacks for hiding,masking and combined protection,and solves the specific problems affecting attack efficiency.The main research contents are as follows:1.To address the problems of noise sensitivity,poor versatility and low efficiency in typical time-domain alignment methods,an alignment method based on trend series dynamic matching is proposed.This thesis analyzes the working characteristics of time-dimension hidding,proposes a piecewise linear representation method based on angle key points,extracts the trend information of power trace,transforms the elements involved in the alignment operation,and designs a dynamic time warping based on cosine distance.Finally,the trend series dynamic matching scheme is formed.Compared with typical alignment methods,this method has better alignment effect and strong anti-noise;and the number of power traces required for the success rate of attack to reach 100% is respectively reduced by 23.8%,24.2%,and 11.3%.2.To address the problem of blinding neighborhood setting in Locally Linear EmbeddingCorrelation Power Attack,an optimized method is proposed.This thesis analyzes the local distribution characteristics of the power traces,proposes the neighborhood adjustment coefficient manifold curvature and local density,and designes the neighborhood dynamic adjusted factor.Combining existing methods,the optimized method is formed.Verified on the simulated power traces with first-order and second-order mask and the public data set with RSM,the attack correlation of this optimized method is stronger,and the attack success rate is respectively increased by 12.9%,43.5%,13.3%.3.To address the problem of low accuracy,poor robustness and slow convergence speed of typical neural network models,an attack method based on the CNN-MGU neural network model is proposed.This thesis analyzes the characteristics of typical neural network models,uses the intermediate value of cryptographic algorithm operations as data labels,applies the feature extraction ability and translation invariance of CNN layer to exploit key information,and adopts MGU layer to deeply mine temporal interdependence of local key information.Finally,an attack method based on the fusion model is formed.Verified on the public dataset with combined protection,compared with the attack methods based on convolutional neural network and longshort-term memory network,the attack accuracy rate based on CNN-MGU method is increased by 5.6% and 3.4% respectively;When the delay amount increases from 0 to 50 and 100,the attack accuracy rate based on the CNN-MGU method still reaches more than 90%,with stronger robustness and faster convergence speed. |
