Profiled Power-Analysis Attacks by an Efficient Architectural Extension of a CNN Implementation

In a recent line of works, several masking and unmasking AES design have been proposed to secure hardware implementations against power-analysis techniques. Although Machine-learning profiling techniques have been successful in security testing during the last years, evaluation of hardware security still requires improvement because of the growing complexity of leakage models against profiled side-channel attacks. In this paper, we propose an improved profiling method to exploit the power consumption of complex cryptographic functions based on Deep-Learning. In order to learn the 256-class Deep neural network of an AES-128, we build successful Convolutional Neural Networks to break its implementation. It has been shown by our experiments that our model achieved a success rate of $\ge 99$% even with a single trace using Keras library with Tensorflow. For the sake of completeness, we investigate the correct ”key rank” according to the number of traces and as a further performance measure, we use ”recall” metric when attacking the third AES SBox. Our model reaches the key rank $\le 10$ with the recall metric $\ge 0.99$.