Handling the positional and chip-to-chip variations of side-channel monitoring

Side-channel monitoring has the potential to offer a non-intrusive way to assess the behavior of an embedded device by monitoring physical signals such as power consumption or electromagnetic emissions by placing a small measuring probe close to the device. Although studies seem to indicate that it is effective in a laboratory environment with monitoring methods adjusted based on the specific target chip, the reliability of it in real world environments remains unclear due to several challenges such as imprecise probe placement and device variability. This thesis aims to investigate the feasibility of using electromagnetic side-channel monitoring under these conditions in order to investigate the potential of autonomous monitoring. Using the Chipwhisperer Husky platform, Arduino targets, and a robotic arm to emulate an autonomous device, electromagnetic traces were collected from multiple chips and positions. Convolutional Neural Networks were then trained and evaluated on this data to determine which software was executing. The results show that models trained on specific chips can attain some accuracy, but fail to obtain reliable accuracy again unseen chips, even when using data from multiple devices for training. However positional variation in training data improved the accuracy to random probe placement. The findings suggest that side-channel monitoring is possible even with lack of accuracy in position, the variation between differing chips pose a large obstacle. Practical uses of the technology might therefore require chip-specific training data with positional variation.