acceptodds
Under review as a conference paper at ICLR 2027

Risk-Aware Gradient Allocation for Spiking Neural Network

Abstract

Surrogate-gradient learning enables the direct training of spiking neural networks (SNNs), but conventional threshold-centered surrogates concentrate backward gradient energy on the membrane states most susceptible to spike flipping. We recast surrogate design as a risk-aware gradient allocation problem and introduce Gradient Risk Exposure (GRE) to quantify this mismatch between gradient concentration and state instability. By minimizing GRE subject to a Kullback–Leibler penalty on deviations from the baseline allocation, we derive the Risk-Calibrated Surrogate Gradient (RCSG), a closed-form exponential calibration applicable to arbitrary baseline surrogates. RCSG redistributes gradient energy toward more stable membrane states without altering the forward neuronal dynamics or requiring an auxiliary robustness loss. On three static-image classification benchmarks, RCSG improves adversarial robustness while maintaining or improving clean accuracy, and its gains hold consistently across surrogate families, network architectures, and attack gradient estimators.

Primary area:
learning theory
Keywords:
Spiking Neural Networks, Surrogate Gradient Learning, Adversarial Robustness, Risk-Aware Gradient Allocation

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_JDBUOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes the Risk-Calibrated Surrogate Gradient (RCSG) to enhance the adversarial robustness of SNNs by reducing Gadient Risk Exposure (GRE) to membrane states that are vulnerable to spike flipping.

Critical strengths

(1) Experimental results on CIFAR-10, CIFAR-100, and Tiny-ImageNet demonstrate the effectiveness of the proposed method in improving adversarial robustness.

(2) The paper provides a relatively comprehensive ablation study to investigate the contributions of different components of RCSG, which helps validate the effectiveness of the proposed RCSG.

Critical weaknesses

(1). In my opinion, conventional surrogate gradients are designed to approximate the derivative of the Heaviside step function, which is a Dirac delta function in the distributional sense. However, RCSG suppresses the gradient near the firing threshold (z=0z=0), where the Dirac delta is concentrated. This design appears to contradict the original motivation of surrogate-gradient learning. Intuitively, such a design may introduce a greater mismatch between the surrogate gradients used in backpropagation and the actual forward spiking dynamics.

(2). The paper is hard to follow because of heavy notations and formulations.

Questions for authors

(1). What is the surrogate gradient used when testing attack methods? Did the authors evaluate multiple attack-time surrogate gradients and report the worst-case robust accuracy, or did they simply select one independent surrogate gradient? Furthermore, several important experimental details are not clearly specified in the paper, such as the number of PGD iterations.

(2) Figure 3(b) presents the clean–robust accuracy trade-off under different calibration strengths. However, the corresponding hyperparameter values are not provided, making it difficult to interpret the observed trends. In particular, RCSG exhibits a non-monotonic relationship between clean and robust accuracy. Could the authors explain the underlying reasons for this non-monotonic behavior?

Reviewer_i92kOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Risk-Calibrated Surrogate Gradient (RCSG) to improve the adversarial robustness of spiking neural networks. It introduces Gradient Risk Exposure (GRE) to quantify the concentration of gradient energy in unstable membrane states and derives a closed-form risk-aware gradient calibration through KL-regularized optimization. Experiments on three image-classification benchmarks demonstrate improved adversarial robustness while maintaining competitive clean accuracy.

Critical strengths
  1. The risk-aware gradient allocation formulation is well motivated and leads to a closed-form solution which is suitable for different surrogate functions.
  2. Ablation studies and evaluations across different architectures and attack estimators show that RCSG improves adversarial robustness across three benchmarks.
Critical weaknesses
  1. The stronger attacks in Table 3 are evaluated only against the conventional surrogate-gradient baseline, which does not include TGO and MPD-SGR.
  2. The Gaussian membrane-perturbation model is not directly validated against input-space adversarial perturbations. The selection and implementation of key calibration parameters also require clarification, such as risk scale (σ\sigma) and calibration strength (τ\tau).
Questions for authors
  1. Several baselines in Table 2, including Dspike, LSG, MPD-AGL, A2SG, MPD-SGR, and TGO, exhibit substantially lower clean accuracy than reported in their original papers. While differences in architectures and experimental settings may contribute to these discrepancies, could the authors explain the performance gaps and clarify how these baselines were reproduced and tuned?
  2. Section 4.1 states that three training regimes are considered: standard training, FGSM adversarial training, and PGD adversarial training. However, Section 4.4 explicitly reports results under natural training, and the adversarial-training results are not presented. Could the authors clarify whether these experiments were conducted and provide the corresponding results?
  3. There are some typos. For example, "Threshod" in Figure 1 should be "Threshold".
  4. The error bars mentioned in Section 4.2 are not clearly visible in Figure 2(b).
Reviewer_Zti6OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The authors propose RCSG (RISK-CALIBRATED SURROGATE GRADIENT), a method that redistributes surrogate gradients during SNN backpropagation from membrane potentials near the firing threshold to those slightly farther away, with the aim of improving adversarial robustness.

Critical strengths
  1. The method is reasonably motivated.
  2. The reported experimental results are promising.
Critical weaknesses
  1. The methodological novelty is limited.
  2. The proposed mechanism requires stronger justification.
  3. The experimental evaluation is limited and omits several essential robustness assessments.
Questions for authors
  1. Although the overall motivation behind RCSG is reasonable, its novelty appears limited. For example, TGO (ICLR 2026) mentioned in the manuscript explores a related idea: improving robustness by discouraging membrane potentials from concentrating near the firing threshold. RCSG instead reduces surrogate-gradient magnitudes near the threshold. However, the causal link between these two effects is insufficiently substantiated. Reducing gradient magnitudes near the threshold does not, by itself, demonstrate that the learned membrane-potential distribution will shift away from it. More direct theoretical justification or empirical evidence is needed to support this connection.
  2. The presentation of the method is difficult to follow, and several symbols are insufficiently defined. Important methodological details are also missing. For example, it is unclear which component the normalization factor C_\tau is applied to and at what stage of the computation. The authors should provide additional implementation details or complete pseudocode to improve clarity and reproducibility.
  3. The manuscript would benefit from clearer writing, better organization, and more precise technical exposition throughout.
  4. The experimental evaluation is limited in scope. In particular, the absence of experiments on DVS datasets leaves the applicability of the method to event-based data insufficiently explored.
  5. Evaluations against black-box attacks are missing and should be included to provide a more comprehensive assessment of adversarial robustness.
  6. The evaluation lacks diagnostic tests to rule out gradient masking or obfuscated gradients (Athalye, Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples, ICML 2018). Because RCSG modifies surrogate gradients to improve robustness, these checks are particularly important for determining whether the reported gains reflect genuine robustness or limitations of gradient-based attacks. Together with black-box attack evaluations, such analyses are essential for substantiating the robustness claims.
  7. What additional training overhead does RCSG introduce? Please quantify its computational cost and training time relative to the baseline.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.