Research On Power Analysis Attack Based On Machine Learning

Although the cryptographic algorithm is theoretically secure,the power consumption information related to the secret key is still leaked during the execution on the physical device.The power analysis attack method refers to that the attacker collects power leakage information and combines the structural characteristics of the cryptographic algorithm itself to analyze it using statistical methods to recover the correct key of the cryptographic algorithm.With the popularity of intelligent hardware devices embedded with cryptographic algorithms and the rapid development of Internet of Things,PA attack poses a serious threat to the security of cryptographic devices.In the past decade,machine learning(ML)has developed rapidly,especially deep learning has demonstrated advantages over traditional technologies in many fields such as face recognition,speech recognition,autonomous driving,etc.Researchers have developed a brand new research direction —— ML-based PA attack,which breaks through the shortcomings of the existing attack methods based on statistics,and makes PA attack more practical,automatic and intelligent.Now,the research on ML-based PA attack is still in the exploratory stage,and many challenging issues need to be further explored.This paper focuses on the security evaluation of the cryptographic chips.Taking the AES algorithm as an example,we conducted a comprehensive and in-depth study of ML-based PA attack.We consider PA attack as a classification task in supervised learning,and give a ML-based PA attack framework that relaxes the assumption of multi-Gaussian distribution of leakage information.Moreover,the ML-based PA attack framework overcomes the limitations of traditional PA attack methods,and is particularly suitable for scenarios where it is difficult to construct an approximate leakage model or extract effective features.On this basis,the research results obtained in this paper are as follows:1.Considering the adaptive analysis of wavelet function and the kernel function of SVM,we propose the Wavelet SVM-based PA attack method.Compared with SVM-based PA attack using other kernel functions,our proposed method has a higher success rate of mask recovery and better efficiency of key recovery.In the mask recovery phase,the success rate of Wavelet SVM is 5% ? 8% higher than SVM-RBF,but the training time is reduced by 30% ? 40%;In the secret key recovery phase,the attack efficiency of Wavelet SVM is 35% ? 50% higher than SVM-RBF.Therefore,Wavelet SVM-based PA attack is very suitable to solve the key recovery problem of cryptographic algorithm(without or with countermeasure)in the small-scale power traces dataset.2.The CNN-based PA attack does not need the prior knowledge of cryptographic implementation and complex feature engineering,which is impossible for the traditional ML-based PA attack.Therefore,CNN-based PA attack can directly executes attack on the cryptographic algorithm with countermeasure(such as mask,clock disturbance,random delay,etc.).Our proposed Deep SCA network model successfully recovers the correct secret key for AES algorithm with countermeasure,and greatly improves the attack efficiency.Compared with the previous state-of-the-art CNN architecture,the trainable parameters of Deep SCA-FC has decreased by 3.9×,and the attack efficiency is up 5× ? 8×;the trainable parameters of Deep SCA-GAP has decreased by 16.4×,and the efficiency is up 1.25× ? 4×.3.In order to significantly improve the impact of class imbalance on ML-based PA attack,we propose the Tri-SMOTE oversampling method.This method incorporates the consideration of the density of the nearest neighbor regions when synthesizing new minority samples,and solves the problem of the sparse distribution of SMOTE for the minority class.Compared with the original class imbalance scenario,Tri-SMOTE improves the attack efficiency of SVM and CNN by more than 30%;the Tri-SMOTE +Tomek links method improves the attack efficiency of SVM and CNN by approximately1×.Therefore,the Tri-SMOTE oversampling method significantly improves the attack efficiency of ML-based PA attack under the class imbalance scenario.