Paper 2020/952

On What to Learn: Train or Adapt a Deeply Learned Profile?

Christophe Genevey-Metat, Benoît Gérard, and Annelie Heuser

Abstract

In recent years, many papers have shown that deep learning can be beneficial for profiled side-channel analysis. However, in order to obtain good performances with deep learning, an attacker needs a lot of data for training. The training data should be as similar as possible to the data that will be obtained during the attack, a condition that may not be easily met in real-world scenarios. It is thus of interest to analyse different scenarios where the attack makes use of ``imperfect" training data. The typical situation in side-channel is that the attacker has access to an unlabelled dataset of measurements from the target device (obtained with the key he actually wants to recover) and, depending on the context, he may also take profit of a labelled dataset (say profiling data) obtained on the same device (with known or chosen key(s)). In this paper, we extend the attacker models and investigate the situation where an attacker additionally has access to a neural network that has been pre-trained on some other dataset not fully corresponding to the attack one. The attacker can then either directly use the pre-trained network to attack, or if profiling data is available, train a new network, or adapt a pre-trained one using transfer learning. We made many experiments to compare the attack metrics obtained in both cases on various setups (different probe positions, channels, devices, size of datasets). Our results show that in many cases, a lack of training data can be counterbalanced by additional "imperfect" data coming from another setup.