In recent years, many papers have shown that deep learning can be beneficial for profiled side-channel analysis. However, to obtain good performance with deep learning, an evaluator or an attacker face the issue of data. Due to the context, he might be limited in the amount of data for training. This can be mitigated with classical Machine Learning (ML) techniques such as data augmentation. However, these mitigation techniques lead to a significant increase in the training time; first, by augmenting the data and second, by increasing the time to perform the learning of the neural network. Recently, weight initialization techniques using specific probability distributions have shown some impact on the training performances in sidechannel analysis. In this work, we investigate the advantage of using weights initialized from a previous training of a network in some different contexts. The idea behind this is that different side-channel attacks share common points in the sense that part of the network has to understand the link between power/electromagnetic signals and the corresponding intermediate variable. This approach is known as Transfer Learning (TL) in the Deep Learning (DL) literature and has shown its usefulness in various domains. We present various experiments showing the relevance and advantage of starting with a pretrained model In our scenarios, pretrained models are trained on different probe positions/channels/chips. Using TL, we obtain better accuracy and/or training speed for a fixed amount of training data from the target device.

Neural Networks (NN) have been built to solve universal function approximation problems. Some architectures as Convolutional Neural Networks (CNN) are dedicated to classification in the context of image distortion. They have naturally been considered in the community to perform side-channel attacks showing reasonably good results on trace sets exposing time misalignment. However, even in settings where these timing distortions are not present, NN have produced better results than legacy attacks. Recently in TCHES 2020, auto-encoders have been used as preprocessing for noise reduction. The main idea is to train an auto-encoder using as inputs noisy traces and less noisy traces so that the auto-encoder is able to remove part of the noise in the attack dataset. We propose to extend this idea of using NN for pre-processing by not only considering the noise-reduction but to translate data between two side-channel domains. In a nutshell, clean (or less noisy) traces may not be available to an attacker, but similar traces that are easier to attack may be obtainable. Availability of such traces can be leveraged to learn how to translate difficult traces to easy ones to increase attackability.