Novel deep learning-based side-channel attack on different-device

By exploiting advances in deep learning (DL), several DL-based profiled side-channel analyses (SCAs) have been performed. DL-based profiling analysis is a technique that trains the relationship between side-channel information (e.g., electromagnetic leaks and power consumption) and the intermediate values of side-channel information on a neural network (NN) and then finds the secret key of an attack device using the trained NN. Recently, cross-device profiled SCA has been proposed to consider the realistic DL-based profiled SCA scenarios. However, it has a limitation in that attack performance is lowered if the profiling and attack devices have different target chips. In this study, an environment in which the attack device has a target chip different from the profiling device is defined as different-device and a novel DL-based profiled SCA on different-device is proposed. In addition, multiscale convolution NN, an NN architecture that trains preprocessed and original data in batches, is used to extract and train the characteristics of each data. We experimented with six boards to verify the attack performance of the proposed method. As a result, applying the proposed method reduced the number of attack traces required for the guessing entropy to converge to zero by up to 45 times compared with without applying it.