Trace Augmentation: What Can Be Done Even Before Preprocessing in a Profiled SCA?

Abstract

Preprocessing is an important first step in side-channel attacks, especially for template attacks. Typical processing techniques, such as Principal Component Analysis (PCA) and Singular Spectrum Analysis (SSA), mainly aim to reduce noise and/or extract useful information from raw data, and they are barely robust to tolerate differences between profiling and target traces. In this paper, we propose an efficient and easy-to-implement approach to preprocessing by applying the data augmentation method from deep learning, whose appropriate parameters can be efficiently determined using a simple validation. Our trace augmentation method, when added prior to existing profiling methods, significantly enhances robustness and improves performance of the attacks. Simulation-based experiments show that our approach not only results in a more robust profiling (even show an enhancement to the known robust profilings), but also works well in the ideal scenario (no distortions between profiling and target traces). The results of FPGA-based and software experiments are consistent to the ones of simulation-based counterparts. Thus, we conclude that the proposed augmentation method is an efficient performance-boosting add-on to profiled side-channel attacks in real world.

Appendices

A The Impact of Augmentation Ratio C

Figure 7 shows the impact of augmentation ratio C in trace augmentation, and we can see that it is insignificant to the improvement.

Fig. 7.

The guessing entropies by varying augmentation ratio C in ideal scenario (no misalignment); simulation-based experiment containing 50 leakage points; 100 repetitions (to compute the guessing entropies) and 2000 profiling traces

B Correlation Matrices

‘Vine’ works in this way: off-diagonal values are derived from a beta distribution whose parameters satisfying $\alpha =\beta$, then perform a linear transform of these values to the interval $[-1.0,+1.0]$ (since beta distribution is defined on the interval [0, 1]). Correspondingly, values of correlation matrix are controlled by the single parameter $\beta$—higher $\beta$ value corresponds to the less dependencies among points of each trace.

The correlation matrices of varied $\beta$ value are provided as Fig. 8, colored according to correlations, from $[-1.0,+1.0]$. It is observed that correlations among points are enhanced as $\beta$ decreasing.

Fig. 8.

Correlation matrix ($50\,\times \,50$) of each $\beta$ parameter: 0.1, 1, 5, 10