Advanced Side-Channel Analysis of USIMs, Bluetooth SoCs and MCUs

Abstract: The use of deep learning for side channel analysis has shown a lot of success in recent years. Impressive results have been presented by many researchers. However, critics of this approach have voiced concerns about the ad hoc methodologies and bespoke neural network designs used in many presented approaches. This thesis first analyzes a possibility of generalizing the selection of neural network architecture for side channel analysis. Then, it presents a simple model for a multilayer perceptron network that does not need to be altered for different targets. Experiments are conducted on three different data sets; power consumption measurements of USIMs, far-field electromagnetic measurements of a Bluetooth device, and power consumption measurements of dedicated XMega victim boards. For each of these sets a model is presented with equivalent or better than state-of-the art results for secret key recovery. Training and testing are done on separate devices in each case. One of the models achieves a classification accuracy of 94.5% from a single measurement. Furthermore, the target and the training device do not even share the same printed circuit board layout. Another model achieves a 47.4% classification accuracy from measurements captured in a manner that is possible in a real-world attack. The thesis also investigates if three different numerical ways of determining the leakage point in unprotected implementations of AES agree. The tests are applied to all three data sets. Finally the thesis evaluates whether the popular transformer architecture is beneficial for side channel analysis.