acceptodds
Under review as a conference paper at ICLR 2027

Analyzing Mask-Breaking Mechanisms: A Methodology for Deep Learning Side-channel Analysis

Abstract

Side-channel analysis (SCA) of cryptographic implementations remains a persistent threat, necessitating the certification of new products by independent evaluation labs. In recent years, deep learning-based SCA (DLSCA) has become a standard approach for worst-case evaluations, due to its ability to defeat diverse countermeasures. While achieving state-of-the-art performance, this comes at the cost of the evaluator's understanding of the attacks themselves. In this work, we propose a novel methodology based on mechanistic interpretability to understand how neural networks recombine shares to defeat the masking countermeasure. Concretely, to isolate the role of individual layers, we freeze the first few layers of a trained model and retrain the remainder while injecting ground-truth information about specific secret shares. This allows us to determine where exploitable share-related information is represented in the network and how it can be used. To demonstrate our approach, we analyze EffCNN and EstraNet on ASCADv1f and ASCADv1r, and EffCNN on the non-shuffled ESHARD benchmark. We find that injecting clean information about one share reveals which information about the other share remains exploitable by the model, highlighting that the models represent substantially more information about individual shares than they originally exploit. Furthermore, our method confirms that the organization of share-related information within the models follows the physical leakage characteristics, without requiring the manual effort of previous approaches and thereby allowing interpretability analyses to scale to more realistic models. Overall, our work suggests that injecting ground-truth share information is a promising direction for interpreting DLSCA models, going beyond prior approaches by revealing not only where share information is represented but also how much of it remains unexploited.

Primary area:
interpretability and explainable AI
Keywords:
Deep-learning, DL-SCA, Interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ExzcOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper aims to interpret how deep-learning-based side-channel analysis (DLSCA) models process masked cryptographic implementations. It injects ground-truth mask information into intermediate network activations at a target layer, while freezing its preceding layers and retraining its subsequent layers. By comparing different mask injection schemes and analyzing the resulting attack performance and learned representations, the authors aim to identify share-related information retained across network layers. The evaluation results suggest that intermediate representations contain additional exploitable share-related information, and the learned mask embeddings reflect the physical leakage characteristics of the datasets.

Critical strengths
  1. The paper studies an interesting problem of understanding how DLSCA models process masked cryptographic implementations.

  2. The evaluation covers different neural network architectures and SCA datasets, with multiple injection schemes and complementary analyses of attack performance and learned representations.

Critical weaknesses
  1. Mismatch Between Research Objective and Methodology: This paper aims to explain how trained DLSCA models internally extract and recombine masked shares. However, the proposed method injects ground-truth mask information at a target layer and retrains its succeeding layers (with preceding layers frozen). This fundamentally changes the model being analyzed, as the retrained network can learn an entirely different recombination strategy using information unavailable to the original model. As a result, the observed attack performance and learned representations only characterize what a newly trained network can exploit from the frozen activations; they cannot reveal how the original model actually performs mask-breaking. Overall, the paper provides no convincing evidence connecting the mechanisms observed after retraining to the computations performed by the original model.

  2. Unsupported Claims of Unexploited Share Information: The paper claims that improved attack performance after mask injection reveals substantial share-related information that the original models learned but failed to exploit. However, this conclusion does not follow from the experiments. Consider a model that already exploits all the information it can reliably extract from the side-channel traces. Since physical leakage is noisy, the model may still be unable to accurately recover the mask, which limits its attack performance. In this case, providing the exact mask can improve attack performance simply by eliminating uncertainty in mask estimation, even though the original model has already made effective use of the available information. Thus, the observed improvement does not distinguish previously unexploited information from the expected benefit of introducing oracle information. Although the paper includes JFW and RM baselines to control for downstream retraining and additional model capacity, these baselines do not address this fundamental ambiguity.

  3. Questionable Motivation and Practical Security Relevance: The paper motivates its approach by arguing that understanding how DLSCA models perform attacks is important for security evaluations and countermeasures. However, understanding a neural network's internal attack computations is fundamentally different from identifying why a cryptographic implementation leaks sensitive information. Even a precise explanation of how a network combines masked shares through convolutional or attention layers does not necessarily identify the program operations or hardware behaviors that are responsible for the leakage. On the other hand, implementation-level analysis and physical leakage analysis can investigate these leakage sources directly, without requiring an explanation of a particular neural network's internal computations. Given this distinction, the paper needs to clarify how its proposed model-level analysis contributes to identifying, understanding, or mitigating side-channel vulnerabilities in the underlying implementation.

  4. Limited Novelty and Significance for XAI and SCA: The paper's contributions to both explainable AI (XAI) and side-channel analysis remain unclear. From an XAI perspective, the proposed methodology combines existing techniques, and the paper does not show what new interpretability capabilities this combination provides. Further, the experiments also do not reveal new or generalizable insights into how neural networks learn or represent information. From an SCA perspective, the main findings mostly confirm existing knowledge about physical leakage and masked implementations. The paper does not identify previously unknown leakage characteristics or additional insights that could guide countermeasure design. In other words, this paper's novelty appears largely limited to applying a particular combination of existing interpretability techniques to DLSCA models. Particularly, as an ICLR submission, the paper's significance and contribution to the broader machine learning community should be clearly justified.

  5. Unsupported Characterization of DLSCA as Worst-Case SCA: This paper characterizes DLSCA as a standard approach for worst-case SCA evaluations. In practice, however, DLSCA's performance depends heavily on model expressiveness. For example, a model with insufficient expressiveness to capture dependencies between masked shares may fail to recover sensitive information, even when the same leakage can be successfully exploited by other SCA techniques. In such cases, the failure of DLSCA reflects limitations of the model and may substantially underestimate the actual SCA threat.

Questions for authors
  1. What evidence shows that the recombination mechanisms learned after mask injection and downstream retraining correspond to the actual computations performed by the original DLSCA models?

  2. How do the experiments distinguish previously unexploited share information from the expected performance improvements resulting from the provided ground-truth mask information?

  3. How does understanding the internal computations of a DLSCA model help identify or mitigate the underlying leakage sources in a cryptographic implementation?

  4. What new interpretability capabilities or generalizable insights does this paper provide to XAI, and what previously unknown SCA knowledge does this paper uncover?

Reviewer_iW5COpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors propose a mechanistic interpretability technique for deep learning side-channel attacks on masked cryptographic hardware. They inject functions of the mask variable into early layers of the models, retrain the later layers, and look at how performance changes/visualize the mask embeddings to draw insights into what information various layers have about the two shares.

Critical strengths
  • The paper explores an important problem: it's poorly understood what information DLSCA attacks exploit and what they leave on the table, which has implications for how they should be used in security evaluations.
  • They explore lots of experimental axes, the manuscript seems to contain quite a bit of raw information.
Critical weaknesses
  1. The links between evidence and core claims aren't very clear to me, and some important details seem to be missing:
    1. A core claim seems to be that injecting the mask while leaving early layers frozen improves performance, therefore the frozen representations retain complementary information about the masked intermediate value. The paper doesn't precisely state what this means or why it should be true. The model must have to predict both mm and m⊕ym \oplus y to predict yy, no? And its overall performance should go up if it gets better at predicting either of the shares? Maybe I'm missing something, but I don't see how this tells us anything interesting about the frozen representations?
    2. There's no pseudocode for the algorithm and code doesn't appear to be released. I think I more or less parsed the approach from the prose, but pseudocode would be easier and less-ambiguous.
    3. The figures are too busy, jargon-heavy, and it isn't clear what trends I should be seeing. They seem to be listing out lots of different things the authors tried rather than supporting specific claims. I recommend the authors tighten these up and defer redundant/uninteresting results to the appendix, simplify the jargon and make the figure + caption self-contained, and explicitly state in the caption what trend/phenomenon I should be observing.
    4. There are missing training details, e.g. learning rate, LR schedule, how many epochs for the training + retraining phases. The authors state that they follow baseline settings from corresponding papers, but their setup seems to deviate significantly from these papers -- e.g. the EstraNet paper attacks a 20k-timestep window of ASCADv1f but the authors state that they attack a 700-timestep window. And this paper does a retraining phase w/ mask injection that the corresponding papers do not have.
  2. For the ASCADv1 datasets the authors seem to consider only one mask/masked sensitive variable. These datasets actually have at least three different mask/masked sensitive variable pairs which significantly leak this sensitive variable (Egger et al, 2022, Fig. 3). Gammell et al. (2026) observe that standard supervised models seem to mostly exploit the pair of variables (ptxt⊕k⊕rin,,rin)(ptxt \oplus k \oplus r_{in},, r_{in}) (notation of Egger et al, 2022), rather than the pair considered in the paper. Can the authors comment on how to interpret their results in light of this, and whether the proposed method is appropriate in settings such as this with multiple distinct masks?
Questions for authors
  1. Why do you learn a single KK when you try the different leakage models? It seems like the same W2V approach should still work -- you could just learn one KK matrix for each of the 9 Hamming weight values or 2 LSB values. I don't see why W2V approach vs. learning a single KK should depend on which leakage model you are using, these seem like orthogonal experimental axes.
  2. Which distributions are you computing the KL divergence between in Fig. 1? Is it the softmax outputs for the baseline model vs. model w/ mask injection after seeing a single trace, averaged over a dataset? Or the distributions you get after accumulating lots of attack traces? Why is this a quantity we care about?
  3. Can your method scale to attacks on uncropped traces e.g. [GPAM] or the 10k--20k-timestep traces considered in the EstraNet paper? It seems like it should, and showing this would strengthen the claim that you scale to "more realistic models".
  4. How concretely might someone use these techniques in practice to mitigate side-channel attacks?
  5. How does your approach differs from existing mechanistic interpretability techniques which have been used for vision/language models, etc.?

[GPAM] Bursztein, Elie, et al. "Generalized power attacks against crypto hardware using long-range deep learning." arXiv preprint arXiv:2306.07249 (2023).

Reviewer_LKH8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work introduces a mechanistic interpretability method for studying how deep learning models combine secret-share information to defeat masking in side-channel analysis. By freezing early layers and retraining the remaining network with ground-truth information about selected shares, the approach reveals where share-related information is represented and how much remains unexploited. Evaluations on ASCADv1 and ESHARD show that the models encode more share information than they initially use, and that its organization reflects physical leakage characteristics, offering a scalable alternative to manual analysis.

Critical strengths

(1) The paper investigates an interesting and relevant topic in side-channel analysis, examining how deep learning models use share-related information to overcome masking.

(2) The author seems to conduct extensive experiments across multiple datasets and model architectures, providing useful evidence to support their findings.

Critical weaknesses

(1) The paper’s writing needs significant improvement, particularly in readability and clarity, beyond correcting typos and grammatical errors. Even after reading the full paper, it is difficult to identify the main contribution and understand the motivation for key design choices, such as the neural network architectures selected for side-channel analysis.

(2) The paper does not clearly define its threat model, making it difficult to understand the assumptions underlying the proposed analysis. For example, does the attacker assume a fixed or randomly generated AES key? Is the attack profiled or non-profiled, and does it require full access to a clone device? If clone-device access is not assumed, how is the deep learning model trained and then applied to the target device to recover the secret key?

(3) The experimental section does not report the method’s computational cost, making it difficult to assess its practical feasibility. The evaluation also appears limited to a small set of datasets; the authors should consider including another widely used side-channel evaluation dataset, such as CHES_CTF, or explain why it is not suitable for this analysis.

Questions for authors

(1) What is the paper’s main contribution?

(2) What threat model does the paper assume?

(3) What is the computational cost of the proposed method?

In addition to addressing these questions, please refer to my comments for further concerns about the paper.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.