The Research On The Technologies Of Electromagnetic Attack Oriented To Cryptographic Chips Based On Machine Learning

Machine learning has become one of the most active and potential technology in the field of the computer. It has been successfully applied in face detection, speech recognition and other aspects. In recent years, some domestic and foreign scholars have introduced machine learning methods to cryptography design and analysis area into which is injected new vitality. At the same time, it has opened up a new direction for a comprehensive cross development for cryptography and machine learning--side channel attack based on machine learning.The security of information system is decided by the weakest component, which may make the cryptographic algorithm insecure due to improper physical implementation even if it is mathematics secure by itself. Side-channel attacks are effective against cipher chips which are cryptographic algorithm execution carriers. It uses a variety of physical information leaked by cipher chip in operation process (such as power, electromagnetic radiation, visible light, sound, etc.) to crack the cryptographic system. Compared with the traditional cryptanalysis method, side channel attacks have smaller key search space and better efficiency. The machine learning can continue to improve and enhance their performance through experience like human beings. So side-channel attacks based on machine learning are important to improve the reliability and automation level. Current research work in this field is still small and many interesting issues need to be further explored.Compared to other side-channel attack methods, electromagnetic attack don’t need to establish direct electrical connection during the attack process which makes it difficult to dected and has strong concealment. In this paper, with the commonly used cryptographic algorithms being targeted, the machine learning method is applied to the side channel attacks by detect electromagnetic radiation. Some important results were achieved:(1) For the shortcoming of plaintext or ciphertext must be known in the traditional template attack, a single bit electromagnetic template attack method aiming at key is proposed. The method can not only reduce the number of templates and he computational complexity, but also use more training data in template creation. However it has low classification accuracy because of no consideration the nfluence of plaintext. To improve effective, the attack area is recommended to locate the key expansion stage. By experimenting on the DES cryptographic algorithm implemented on the MCU’s, the results show that the method can infer exactly the right key just only use a sample curve without knowing the plaintext or ciphertext.(2) To avoid the the problem of "sick" covariance matrix and floating-point overflow, a fast template attack method based on polynomial simplification and transformation is proposed by keeping the rank of template matching probabilities unchanged. The method uses public covariance matrix instead of the covariance matrix which makes it not only relax the restrictions of the covariance matrix being reversible, but also improve the success rate of attacks while reducing the computational complexity because of similarity to the real covariance matrix. By comparative experiments on the covariance matrix, the public covariance matrix and identity matrix, the result proved the method effective.(3) For the common problem of high dimension and small sample in supervised learning, at first a new dimension reduction algorithm shorted for SPP is proposed to combine with the advantages of feature selection and feature extraction algorithm. Secondly, a new multi-class SVM classifier which only needs to train K-1binary SVMs instead of (K-1) K/2is proposed based on the hamming model with strict order. Through experiments on8-bit microcontroller which is implemented DES software, the results show that SPP method is better than the PCA and multi-class SVM attack is better than the template attack.(4) Supervised learning requires a large number of labeled training data and in actual situation it is difficult to get such complete data set. An unsupervised attack method is proposed. The method constructs linear regression model of cryptographic device at first. Then the parameters of the model are calculated by the method of least squares. To quantify the estimation error, the goodness of fit model is used and the correlation coefficient of determination is computed for each sample point. To eliminate interference of ghost peaks, a modification method based on normalized inter-class variance is putted forward. The experimental results show the effectiveness of the method.