Improving cross-device attacks using zero-mean unit-variance normalization

Template attacks are a very powerful form of side-channel analysis. It is assumed an adversary has access to a training device, identical to the device under attack, to build a precise multivariate characterization of the side-channel emissions. The training and test devices are assumed to have identical, or at least very similar, electromagnetic emissions. Often, when evaluating the effectiveness of a template attack, training and test data are from the same-device. The effectiveness of collecting training and test data from different devices, or cross-device attacks, are evaluated here using 40 PIC microcontroller devices. When the standard template attack methodology fails to produce adequate results, each step is evaluated to identify device-dependent variations. A simple pre-processing technique, normalizing the trace means and variances from the training and test devices, is evaluated for various test data set sizes. This step improves the success key-byte extraction rate for same part number cross-device template attacks from 65.1 to 100 % and improves attacks against similar devices in the same-device family. Additionally, it is demonstrated that due to differences in device leakage, minimizing the number of distinguishing features reduces the effectiveness of cross-device attacks.