Side-channel attacks on elliptic curve cryptosystems based on machine learning techniques

This thesis presents a promising approach to Side-channel attacks on the cryptosystems based on elliptic curve cryptography (ECC). This approach is based on machine learning analysis in characterisation of side-channel information. The original contribution of this thesis is to verify the performance of machine-learning techniques in terms of neural networks (NN), support vector machines (SVM) and principal component analysis (PCA). In this project, PCA is used as a powerful algorithm in the preprocessing stage to decrease the computational complexity of the input dataset, while SVM and NN are utilised as efficient multi-class classiffiers to recognise and classify different patterns of side-channel information. In order to investigate the proposed method, an experiment based on the power consumption and electromagnetic emission of an field-programmable gate array (FPGA) implementation of ECC was conducted. Regarding our experimental results based on an FPGA implementation of ECC, PCA can be used as a strong preprocessing stage to reduce the signal-noise ratio, data-set dimension and algorithm complexity. In addition, after verifying the performance of different techniques and specifications such as kernel functions, neural-network architect and parameters, we inferred that the most efficient machine-learning techniques for side-channel information characterisation are LVQ neural network (with a number of hidden layers between 90 and 100), and SVM with Gaussian RBF kernel function with parameter p value of 5 and 50 for CS and M-SVM2 SVM models respectively with about 80 to 85 % accuracy.