Templates vs. Stochastic Methods

Abstract

Abstract. Template Attacks and the Stochastic Model provide advanced methods for side channel cryptanalysis that make use of 'a-priori' knowledge gained from a profiling step. For a systematic comparison of Template Attacks and the Stochastic Model, we use two sets of measurement data that originate from two different microcontrollers and setups. Our main contribution is to capture performance aspects against crucial parameters such as the number of measurements available during profiling and classification. Moreover, optimization techniques are evaluated for both methods under consideration. Especially for a low number of measurements and noisy samples, the use of a T-Test based algorithm for the choice of relevant instants can lead to significant performance gains. As a main result, T-Test based Templates are the method of choice if a high number of samples is available for profiling. However, in case of a low number of samples for profiling, stochastic methods are an alternative and can reach superior efficiency both in terms of profiling and classification.

Profiling DPA: Efficacy and Efficiency Trade-Offs

“…These are obtained by averaging over 1,000 repeat experiments on randomly drawn balanced samples (i.e. comprising an equal number of replicates per intermediate value) with Gaussian noise at high (8), medium (1) and low (0.125) signal-to-noise ratios (SNRs) 5 as model degree ranges from 1 through to 8. 6 Criterion 3 can be assessed straightforwardly by computing theoretic distinguishing vectors for correlation DPA using the asymptotically fitted model cor-4 'Power' in this context refers to statistical power and should not be confused with the 'Power' in DPA.…”

“…6 Criterion 3 can be assessed straightforwardly by computing theoretic distinguishing vectors for correlation DPA using the asymptotically fitted model cor-4 'Power' in this context refers to statistical power and should not be confused with the 'Power' in DPA. 5 We define the SNR as var(L(V )) var(ε) , where L is the data-dependent leakage function (the variance of which is computed with respect to the distribution of the intermediate value V , which is uniform throughout in our analysis) and ε is the independent noise. 6 To reduce computational complexity we take the usual strategy (see, for example, [1])…”

“…This is a popular context for research as it is both highly realistic (e.g., frequently observed in devices built using CMOS logic) and straightforward to analyse. Indeed, many previous works evaluating profiling methods [5,13] …”

“…Thus the theoretic distinguishing vectors for the S-box attacks have larger nearest-rival margins and the corresponding sample sizes are smaller than those for the attacks against AddRoundKey. It is clear that this straightforward leakage scenario-which is the one investigated in [5] and [13]-lends itself very naturally to linear regression-based profiling, as the true data-dependent leakage function can be easily and precisely approximated with only linear terms. Our experiments indicate that the profiling stage requires around thirty times fewer training samples than 'classical' templates with no trade-off on model precision, fit, nor DPA performance.…”

“…Of course, as soon as more than one profiling method exists the natural question to ask is which is 'better' in practice? Previous studies evaluating linear regression relative to 'classical' templates [5,11,13] have demonstrated substantial efficiency gains in some typical leakage scenarios.…”