Gambling for Success: The Lottery Ticket Hypothesis in Deep Learning-based SCA

Guilherme Perin, Lichao Wu, Stjepan Picek

Introduction

Side-Channel Analysis (SCA) has long been a critical field in hardware security, where attackers exploit physical leakages—such as power consumption or electromagnetic emissions—to extract secret keys from cryptographic implementations. While Deep Learning (DL) has significantly advanced the efficacy of these attacks, it has introduced a challenging "Goldilocks" problem. Small neural networks are often too simple to capture complex leakage patterns, while large, overparameterized networks are prone to overfitting and require exhaustive hyperparameter tuning.

This paper explores a solution to this trade-off by applying the Lottery Ticket Hypothesis (LTH) to the SCA domain. The hypothesis suggests that within large neural networks, there exist tiny sub-networks ("winning tickets") that, when trained in isolation, can match or even exceed the performance of the original model. By identifying these tickets, researchers can achieve superior attack performance with significantly reduced computational overhead.

The Lottery Ticket Hypothesis in SCA

The researchers implement a systematic "prune and reset" workflow to identify these high-performing sub-networks. Unlike standard pruning techniques that simply aim for compression, the LTH approach emphasizes the importance of the initial weight configuration.

The Lottery Ticket Hypothesis workflow: training a baseline, pruning small weights, and re-initializing to identify winning tickets.

The Prune and Reset Workflow

  1. Train a Baseline: The process begins by training a relatively large, overparameterized baseline model (typically an MLP or CNN).
  2. Pruning: The "weakest" connections—those with the smallest weight magnitudes—are identified and removed. The study evaluated sparsity levels ranging from 1% to 99%.
  3. The "Winning Ticket" Reset: This is the critical step. Instead of randomly re-initializing the remaining weights, they are reset to the exact values they held at the start of the baseline training.
  4. Retraining: This slimmed-down sub-network is then retrained.

In the context of SCA, success is measured by Guessing Entropy (GE), which quantifies the number of traces an attacker requires to successfully recover a secret key, rather than standard classification accuracy.

Experimental Setup and Datasets

The methodology was validated using three standard benchmarks for AES-128 implementations, representing varying levels of complexity and noise:

  • ASCAD (Fixed Key): Data from an 8-bit AVR microcontroller using a fixed secret key.
  • ASCAD (Random Keys): A more difficult variant using variable keys to test generalization.
  • CHES CTF 2018: A highly complex dataset featuring masked AES-128 on a 32-bit STM microcontroller, known for significant noise and countermeasures.

Key Insights and Results

The study yielded several significant findings that confirm the relevance of LTH for cryptographic attacks:

  • Existence of Winning Tickets: The research confirms that LTH holds true for side-channel data. Massive models are not a requirement for breaking AES; rather, a specific, small subset of parameters is responsible for the model's success.
  • Pruning as a Regularizer: Surprisingly, pruned "winning tickets" often outperformed the full baseline models. Pruning appears to act as a powerful regularization tool, stripping away the noise that typically causes large models to overfit to specific trace artifacts.
  • The Criticality of Initialization: Comparing LTH initialization (resetting to original values) against random initialization of the same pruned structure showed that LTH was consistently more stable. This proves that the initial random "luck" of weight distribution is a primary driver of a model's eventual success.
  • Efficiency and Compression: The researchers successfully recovered keys using networks where 90% to 95% of the weights were removed. On the challenging CHES CTF 2018 dataset, while baseline models often failed to converge, the LTH-pruned versions successfully reached a Guessing Entropy of 1.

Conclusion and Future Work

By demonstrating that 90% of a network's weights can be pruned without sacrificing attack power, this work paves the way for more efficient SCA deployments on resource-constrained devices. However, several questions remain. Future research may look into more sophisticated iterative pruning strategies or investigate why specific weights are selected as part of a winning ticket to improve the explainability of DL-based SCA. Additionally, testing the hypothesis against more complex cryptographic countermeasures beyond Hamming Weight leakage models remains a vital next step.

Summary updated

BibTeX
@article{perin2021gambling,
  author = {Guilherme Perin and Lichao Wu and Stjepan Picek},
  title = {{Gambling for Success: The Lottery Ticket Hypothesis in Deep Learning-based SCA}},
  journal = {IACR Cryptol. ePrint Arch.},
  year = {2021},
  url = {https://openreview.net/forum?id=7bkDHhxAYV9q},
}
Figure