On the Performance of Multilayer Perceptron in Profiling Side-channel Analysis

Leo Weissbart, Stjepan Picek, Lejla Batina

Introduction

Side-channel analysis (SCA) remains a critical threat to cryptographic hardware, where attackers exploit physical leakages such as power consumption or electromagnetic emissions to extract secret keys. In recent years, the research community has shifted its focus toward Convolutional Neural Networks (CNNs), primarily due to their inherent ability to handle "jitter"—random timing shifts or delays introduced as countermeasures. Consequently, the simpler Multilayer Perceptron (MLP) has often been relegated to a baseline role or dismissed as ineffective against modern, desynchronized datasets.

This paper challenges the prevailing hierarchy by systematically investigating the performance of MLPs. Through a comprehensive hyperparameter exploration, the study demonstrates that well-tuned MLPs are not only viable but surprisingly potent, capable of breaking implementations protected by both masking and random delay countermeasures.

Methodology

To evaluate the sensitivity and robustness of MLPs, the study employs a rigorous grid search approach for hyperparameter optimization. Rather than searching for a single "best" model, the goal is to map the performance landscape across various architectural configurations.

Table 1: Hyperparameter ranges explored in the grid search.

The evaluation utilizes two primary metrics and techniques:

  1. Guessing Entropy (GE): A standard SCA metric measuring the average rank of the correct key among all candidates. A GE of 0 indicates a successful attack.
  2. Integrated Gradients: An explainability technique used to visualize which specific samples within a power or EM trace contribute most to the model's prediction, helping to distinguish between models that have learned cryptographic features and those that have not.

Experiments were conducted on two public benchmarks: ASCAD (software AES-128 with first-order Boolean masking) and AES_RD (AES-128 with random delay interrupts).

Key Insights and Results

Architectural Stability and Trade-offs

A significant finding is the relative stability of MLPs. Unlike some deep learning models that require precise tuning to function, many different MLP configurations successfully converged to a Guessing Entropy of 0. This suggests that researchers may not need exhaustive searches to find working models for SCA tasks. Furthermore, the results support a practical application of the Universal Approximation Theorem: a lack of depth can often be compensated for by increased width, and vice versa, providing flexibility in hardware-constrained environments.

Resilience to Countermeasures

Contrary to the assumption that CNNs are necessary for desynchronized data, MLPs proved highly effective against the AES_RD dataset. All tested architectures reached a GE below 30 within 2,500 traces, with even minimal models (e.g., 1 layer, 200 neurons) succeeding. For the masked ASCAD dataset, many configurations achieved a GE of 0 within 1,000 traces, with a "sweet spot" identified around 6 layers and 200 neurons.

Figure 1: Performance landscape and explainability of MLP architectures. (a) and (b) show the GE landscape for ReLU and Tanh; (c) shows GE convergence across all architectures; (d) illustrates feature importance via Integrated Gradients.

Activation Functions: ReLU vs. Tanh

The choice of activation function significantly impacts the performance landscape. As shown in Figure 1 (a) and (b), ReLU generally achieved the absolute highest performance (lowest GE). However, Tanh exhibited greater stability, providing consistent results across a broader range of layer and neuron counts. This suggests that while ReLU is better for optimization, Tanh might be a safer default for exploratory analysis.

The Role of Feature Selection

Interestingly, the study found that traditional statistical feature selection, such as Difference-of-Means (DoM), often hindered MLP performance. This indicates that MLPs are capable of extracting value from raw, noisy traces and that manual feature reduction might inadvertently discard subtle non-linear relationships essential for breaking complex countermeasures.

Conclusion

This research rehabilitates the Multilayer Perceptron as a powerful tool in the side-channel analyst's arsenal. By demonstrating that MLPs can effectively overcome both masking and random delays with relatively simple architectures, the paper suggests that the complexity of CNNs may not always be required. Future work will explore the portability of these models—training on one device to attack another—where the simplicity of the MLP may offer additional advantages in robustness.

Summary updated

BibTeX
@article{weissbart2019performance,
  author = {Leo Weissbart and Stjepan Picek and Lejla Batina},
  title = {{On the Performance of Multilayer Perceptron in Profiling Side-channel Analysis}},
  journal = {IACR Cryptol. ePrint Arch.},
  year = {2019},
  url = {https://openreview.net/forum?id=Yhd1ayXhAG3},
}
Figure