Résumé
This chapter presents two classes of threat models that are considered in side‐channel attacks, namely,
supervised
and
unsupervised
attacks. Supervised attacks aim at trading off as much physical assumption as possible by leveraging a preliminary physical characterization of the leakage. Gaussian templates are efficient models, as long as the trace dimensionality (i.e. the number of time samples considered in each trace) remains low (typically around D = 10). However, when the dimensionality increases, some computational complexity overheads may appear. It is noticeable that at the extreme of the generative models spectrum, we may find autoencoders, which are unsupervised variants of multi‐layer perceptrons and convolutional neural networks, and even some non‐parametric models, like the kernel density estimator. Those models are quite powerful, in the sense that they merely require any prior bias, but as a drawback they are very prone to the curse of dimensionality.