The EVIL Machine: Encode, Visualize and Interpret the Leakage

Unsupervised side-channel attacks allow extracting secret keys manipulated by cryptographic primitives through leakages of their physical implementations. As opposed to supervised attacks, they do not require a preliminary profiling of the target, constituting a broader threat since they imply weaker assumptions on the adversary model. Their downside is their requirement for some a priori knowledge on the leakage model of the device. On one hand, stochastic attacks such as the Linear Regression Analysis (LRA) allow for a flexible a priori, but are mostly limited to a univariate treatment of the traces. On the other hand, model-based attacks require an explicit formulation of the leakage model but have recently been extended to multidimensional versions allowing to benefit from the potential of Deep Learning (DL) techniques. The EVIL Machine Attack (EMA), introduced in this paper, aims at taking the best of both worlds. Inspired by generative adversarial networks, its architecture is able to recover a representation of the leakage model, which is then turned into a key distinguisher allowing flexible a priori. In addition, state-of-the-art DL techniques require 256 network trainings to conduct the attack. EMA requires only one, scaling down the time complexity of such attacks by a considerable factor. Simulations and real experiments show that EMA is applicable in cases where the adversary has very low knowledge on the leakage model, while significantly reducing the required number of traces compared to a classical LRA. Eventually, a generalization of EMA, able to deal with masked implementation is introduced.