Improving deep learning assistedfar-field electromagnetic sidechannelattacks on AES : Effects on attack efficiency from using additive noise and otherdata augmentation techniques

Abstract: Profiled side-channel attacks on hardware implemented cryptographic algorithms have been a well-researched topic for the past two decades and many countermeasures against these attacks have been proposed and adopted by the industry. Recently, a new form of far field EM side channel called "screaming channel attacks" have been highlighted. The side channel affects mixed-signal chips used in widespread wireless communication protocols, such as Bluetooth and Wi-Fi. In suchsystems, the radio transmitter may unintentionally broadcast sensitive information from hardware cryptographic components or software executing on the CPU due to substrate coupling. Attacks exploiting such a side channel may succeed over a much longer distance than attacks exploiting usual EM side channels. Deep learning assisted attacks on this new side-channel have recently been demonstrated to further improve the attack efficiency (less data needed in order to classify the underlying AES-128 key). These attacks rely on timeseries of electro-magnetic emissions, captured via a low-noise cable, to train the neural network while timeseries captured using anantenna and software-defined radio are used for classification. As with most projects involving machine learning algorithms the question of what regularization strategies to use is a central one. Many strategies exist in order to potentially reduce the test error, avoid overfit and improvegeneralizability, though which strategies is best to use varies between projects. It is well knownthat the addition of noise during preprocessing of data for training neural network classifiers can bring benefits. No previous research has dealt with how to handle additive noise in side-channelattacks like the case previously described. This thesis investigated if the use of realistic modelled additive noise could bring more benefits to improve models’ ability to generalize compared to conventional additive noise (Gaussian noise). Preprocessing methodologies such as feature scaling and denoising were also evaluated and the quality of captured traces assessed using several metrics. The attack efficiency in two environments with different multipath propagation were also investigated. The results show that using realistic noise did not significantly improve models’ ability to generalize compared to conventional additive noise. No denoising method did improve the attack efficiency but attacks in environments with more prevalent multipath propagation did improve attack efficiency. Trace sets with greater mean signal amplitude and aligned trace sets correlated well with improvements in attack efficiency.