Leakage Certification Made Simple

Side channel evaluations benefit from sound characterisations of adversarial leakage models, which are the determining factor for attack success. Two questions are of interest: can we estimate a quantity that captures the ideal adversary (who knows the distributions that are involved in an attack), and can we judge how good one (or several) given leakage models are in relation to the ideal adversary?

Existing work has led to a proliferation of custom quantities (the hypothetical information HI, perceived informatino PI, training information TI, and learnable information LI). These quantities all provide only (loose) bounds for the ideal adversary, they are slow to estimate, convergence guarantees are only for discrete distributions, and they have bias.

Our work shows that none of these quantities is necessary: it is possible to characterise the ideal adversary precisely via the mutual information between the device inputs and the observed side channel traces. We achieve this result by a careful characterisation of the distributions in play. We also put forward a mutual information based approach to leakage certification, with a consistent estimator, and demonstrate via a range of case studies that our approach is simpler, faster, and correct.