Efficient template attacks based on probabilistic multi-class support vector machines

Abstract

Common template attacks are probabilistic relying on the multivariate Gaussian distribution regarding the noise of the device under attack. Though this is a realistic assumption, numerical problems are likely to occur in practice due to evaluation in higher dimensions. To avoid this, a feature selection is applied to identify points in time that contribute most information to an attack. An alternative to common template attacks is to apply machine learning in form of support vector machines (SVMs). Recent works brought out approaches that produce comparable results, respectively better in the presence of noise, but still not optimal in terms of efficiency and performance. In this work we show how to adapt the SVM template approach in order to considerably reduce the effort while carrying out the attack and how to better exploit the side-channel information under the assumption of an attack model with a strict order, e.g. Hamming weight model. © 2013 Springer-Verlag.