acceptodds
Under review as a conference paper at ICLR 2027

Temporal Quotients for Certified Count-Preserving Spike-Retiming Smoothing

Abstract

Moving spike times can change a spiking neural network's prediction even when event counts remain fixed. We study how to certify robustness to such retiming while preserving event counts and distinct timestamps. We introduce temporal quotient smoothing for discretized event streams under a global budget on total spike displacement. A shared random partition maps each block's events on each line to a common configuration with the same count and distinct timestamps. If retiming crosses no sampled boundary, the transformed inputs coincide, yielding a finite-vote certificate for the smoothed classifier. To control the displacement introduced by smoothing, we minimize expected transport for a fixed event source and placement rule under boundary-hit constraints. A compact dual hierarchy and feasible kernels bound the remaining transport improvement; matching bounds establish periodic-kernel optimality for selected uniform fixed-count families. At global displacement budget on the evaluated fixed checkpoints, Period-5 raises certified accuracy over all-or-none smoothing from to on DVS-Gesture with VGGSNN and from to on N-MNIST. A compact DVS classifier gives a small reversal, while source and decoder fitting can lower both transport and certified accuracy. Temporal quotients thus make count-preserving kernel design verifiable, with classification benefits that depend on the source, classifier, and budget.

Primary area:
optimization
Keywords:
Spiking Neural Networks, Spike Retiming, Certified Robustness, Temporal Smoothing, Optimal Transport, Randomized Smoothing

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_2H3HOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes temporal quotient smoothing for certifying robustness to count-preserving spike retiming. It first partitions the timeline into blocks and then within each block replaces the events with a fixed canonical arrangement that has the same count. Meanwhile, it designs the partition distribution to minimize the displacement introduced by smoothing and formulates a partition LP, gives a compact dual hierarchy and a coloring/path-cover construction. Experiments on MNIST, N-MNIST and DVS-Gesture at B=2 compare Period-5 with an AON kernel and other baselines.

Critical strengths
  1. This paper discusses an emerging certification problem for spiking neural networks against spike-retiming attacks.
  2. This paper provides a solid mathematic derivation for the proposed certificate and the compact optimization formulation.
  3. This paper is well motivated, where Standard randomized smoothing does not apply to binary event streams and the block-count quotient solves this simply.
Critical weaknesses
  1. The practical significance of the evaluated attack budget is unclear. The budget of the proposed certification is B=2, while at this budge, the only reported attack on an undefended model flip none of 274 DVS predictions, and no B=2 attack results are given for MNIST or N-MNIST. More comprehensive experiments are needed to establish whether this budget poses a meaningful threat.
  2. The overall writing of this paper makes reader hard to follow. This paper presents lot of mathematics while lacks a clear hierarchy among those propositions and theorems. This paper also utilizes substantial combinatorial-optimization terminology with limited intuition. A structural revision is recommended to prioritize its core methodology such as quotient construction, operational certificate, and kernel design.
  3. The connection between the optimization theory and classification benefits is weak. A large part of the paper studies how to minimize the displacement caused by smoothing. However, the paper’s own experiments show that reducing displacement can lower certified accuracy. AON also outperforms Period-5 on the compact DVS model. The mathematical results may be solid, but their benefit to classification remains unclear.
  4. The experimental settings are limited. This paper evaluates AON, calibrated p4, and a specific DSSN-JCR adaptation, with the latter evaluated primarily on MNIST. These comparisons do not sufficiently establish the proposed method’s advantage over other feasible smoothing designs under the same threat model. Additionally, the N-MNIST improvement is evaluated on a single checkpoint, while the comparison with AON reverses on the compact DVS model.
Questions for authors
  1. Can the author explain the practical significance of certifying robustness with B=2? Could you report a strong attack at the certified budget on all datasets and discuss whether a meaningful budget can be reached.
  2. What is the inference latency and energy overhead compared with the undefended SNN?
Reviewer_BBqJOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes methods for defending against spike train perturbations under a given budget and perturbation model. A key measure is certified accuracy of a spiking neural network, which corresponds to the fraction of inputs classified correctly at total-displacement budget BB. The perturbation model focuses on spike train jitters, where spikes can move, but the spike counts must remain the same.

Critical strengths

The aim of making SNNs robust to spiking train jitter is valuable, as real-world deployments can encounter spike noise, and we would like to ensure reliability of the system.

Critical weaknesses
  • The results are limited to a very particular type of spike train perturbation model, namely spike jittering, and does not address other modes of edits, such as spike deletions, etc., which are of practical significance.
  • The paper is not ready for presentation at ICLR, as written. The introduction would benefit from a re-write focusing on situating the reader in the context of the problem, why it's important, and what the key definitions are (e.g., certification, etc.). The methods section is too mathematically dense and should focus on a few key results you would like to highlight for the reader. The empirical results must be extended to longer spike sequences and larger edit budgets (e.g., T≫10T\gg10, B≫2B\gg2) to provide convincing results.
Questions for authors

Future work

  • Is there a generalization to all of this to continuous times? Do you need discrete time bins?
Reviewer_byPPOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes temporal quotient smoothing to certify spiking neural networks against spike-retiming attacks that preserve event counts. The method partitions the time axis into random blocks and maps spikes within each block to canonical timestamps while preserving per-line event counts and distinct spike times. The authors derive a finite-vote robustness certificate under a global spike-displacement budget. They also formulate a linear program to minimize the displacement introduced by smoothing, develop compact dual bounds, and prove the optimality of periodic kernels for selected uniform spike distributions.

Experiments on MNIST, DVS-Gesture, and N-MNIST compare the proposed kernels with all-or-none smoothing and other baselines. At a displacement budget of B=2B=2, Period-5 improves certified accuracy from 53.8% to 62.8% on DVS-Gesture and from 25.6% to 76.8% on N-MNIST. The results also show that certification gains depend on the classifier and that minimizing spike displacement does not consistently improve certified accuracy.

Critical strengths

1. A rigorous certification framework for count-preserving spike retiming. The paper constructs a temporal projection that preserves per-line spike counts and distinct timestamps, and derives a finite-vote certificate for all valid retimings within a specified global displacement budget (Theorem 3). The guarantee is mathematically precise and does not rely on sampling estimates. This is a useful extension of existing smoothing methods to a constrained event representation.

2. Nontrivial theoretical results with verifiable optimality guarantees. The compact dual formulation (Theorem 5) provides bounds for kernel optimization, while Theorems 7–9 establish exact periodic-kernel optimality for selected uniform fixed-count distributions. The authors also report independent numerical checks against full linear programs. These results give the paper a substantive theoretical contribution, even though their practical value for robust classification remains insufficiently demonstrated.

Critical weaknesses

1. Limited methodological novelty. The certification mechanism closely follows Deterministic Smoothing with Splitting Noise (DSSN; Levine and Feizi, ICML 2021). Both methods use random partitions, canonical projections, and projection agreement to certify predictions. The count-preserving temporal construction and kernel optimization are new extensions, but the underlying certification principle is established. The comparison covers only one DSSN-JCR adaptation, and the related work omits Mukhoty et al. (ICLR 2024), which studied certified SNN robustness under a different threat model. The paper does not adequately establish the significance of its additions over existing smoothing methods.

2. An insufficiently justified attack budget. The main results use a global displacement budget of B=2B=2. Yet Appendix B.11 reports that the official spike-retiming attack changes none of 274 initially correct DVS-Gesture predictions at this budget. At B=3B=3, only one of 1,536 latency-MNIST evaluations produces a base-model prediction change. These results do not prove that stronger attacks would fail, but they provide no evidence of meaningful vulnerability at the budgets emphasized in the paper. Moreover, B=2B=2 limits total displacement across the entire input, making it substantially more restrictive than a per-spike jitter constraint.

3. No demonstrated certification under damaging attacks. At B=2000B=2000, the retiming attack changes nine of eleven base-model predictions, but Period-5 certifies none. Appendix B.11 also states that the fixed Period-5 certificate becomes vacuous for B≥5B\geq5. The authors report valid larger-budget attacks that change Period-5 predictions themselves. Although the eleven-example test does not estimate population-level vulnerability, the paper fails to establish useful certification in the attack regime where the original classifier is demonstrably vulnerable.

4. Substantial accuracy loss and a poorly justified optimization objective. On DVS-Gesture, Period-5 reduces clean accuracy from 95.1% to 85.4%, without demonstrating a corresponding security benefit against damaging attacks. Much of the theory minimizes expected spike displacement, but lower transport does not reliably improve certified accuracy. Appendix B.7 shows that decoder fitting reduces transport by 27.82% while decreasing certified accuracy across all three checkpoints. Classifier-aware selection among 14,721 candidates also produces no confirmed improvement on held-out data. The practical value of optimizing transport remains unclear.

5. Certification gains are not equivalent to improvements in worst-case accuracy. On N-MNIST, Period-5 improves certified accuracy over calibrated p4 from 25.6% to 76.8%, but achieves lower smoothed accuracy (95.3% versus 97.0%). Appendix B.9 shows that all 512 additional correct certificates come from inputs already classified correctly by p4. This is a genuine improvement in certification coverage, but it does not establish a 51.2-point improvement in worst-case adversarial accuracy. The paper provides no systematic adaptive-attack comparison to determine the actual difference.

6. Limited generality and incomplete comparisons. All-or-none smoothing achieves higher certified accuracy than Period-5 on all three Conv–LIF checkpoints, although the paired confidence intervals include zero. The inference costs are also unmatched: Period-5 uses up to ten distinct projections, compared with at most three for all-or-none. The natural-event experiments use ten-bin, binarized representations rather than raw timestamps, and the uniform-K3 experiment samples spike times independently of class labels. The paper also lacks a matched evaluation against existing timing-aware defenses, including the adversarial training studied by Yu et al. ("Time Is All It Takes: Spike-Retiming Attacks on Event-Driven Spiking Neural Networks", ICLR 2026).

Questions for authors
  1. Novelty: What is fundamentally new in the certification mechanism compared with DSSN (ICML 2021), beyond preserving spike counts and adapting the partition construction? How does the method compare with other count-preserving adaptations?

  2. Threat-model relevance: Why is B=2B=2 a meaningful certification budget when the official retiming attack fails on all 274 tested DVS-Gesture examples? How does this global budget relate to realistic per-spike timing perturbations?

  3. Effective defense: At B=2000B=2000, nine of eleven base predictions change, but Period-5 certifies none. Can the proposed framework provide nontrivial certificates at budgets where attacks succeed? How can it overcome the fixed Period-5 certificate's vacuity for B>=5B >= 5?

  4. Accuracy cost: What justifies the 9.7-percentage-point loss in clean accuracy on DVS-Gesture? Can the authors demonstrate a corresponding security benefit under a meaningful attack budget?

  5. Optimization objective: Why minimize expected transport when reducing it by 27.82% decreases certified accuracy across all three checkpoints? Is there a useful theoretical or empirical relationship between transport and certified classification performance?

  6. Actual adversarial robustness: On N-MNIST, p4 has higher smoothed accuracy but substantially lower certified accuracy. How much of Period-5's reported 51.2-point certification gain corresponds to improved worst-case accuracy? Can the authors provide adaptive-attack results against both smoothed classifiers?

  7. Baselines and generality: Why does Period-5 underperform all-or-none on all three Conv–LIF checkpoints? How does it compare with timing-aware adversarial training (Yu et al., ICLR 2026) under matched threat models, clean accuracy, and inference costs?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.