acceptodds
Under review as a conference paper at ICLR 2027

Surrogate Gradients Misallocate Adversarial Sensitivity in Spiking Neural Networks

Abstract

Spiking neural networks (SNNs) offer a promising basis for low-power intelligence through binary events. This discrete dynamic makes backpropagation-based training reliant on surrogate gradients. However, conventional surrogate functions with fixed profiles fail to adopt to decision-relevant neuronal dynamics under adversarial perturbations, limiting the effectiveness of robust learning in SNNs. Through systematic comparative experiments, we reveal how this lack of adaptation manifests as adversarial sensitivity misallocation, directing backward sensitivity away from membrane states involved in perturbation-induced spike flips. We therefore propose Flip-Aware Reallocation (FARE), which derives dynamic response ranges from decision-weighted spike flips. By making surrogate gradients aware of perturbation-induced flips, FARE reallocates backward sensitivity to bring the learning signal into closer alignment with the network's decision-relevant adversarial response. Across static, sequential, event-based, and large-scale tasks, FARE outperforms existing state-of-the-art methods in both adversarial robustness and clean accuracy.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Robustness, Spiking Neural Network, Surrogate gradient

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_4y1WOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper is motivated by the inconsistency between surrogate gradients and the actual forward dynamics of hard-spiking neurons. Guided by decision-relevant spike flips, the proposed FARE method adaptively adjusts the parameters of surrogate gradients and combines the original and adjusted surrogate gradients through averaging. Experiments on both small- and large-scale datasets demonstrate the effectiveness of FARE compared with baseline methods using original surrogate gradients.

Critical strengths

S1. The authors clearly demonstrate the inconsistency between surrogate gradients and hard-spike forward dynamics through extensive toy experiments, providing a clear motivation for using spike flips to address this issue.

S2. The experiments are comprehensive, with extensive analyses conducted to evaluate the effectiveness of the proposed method.

Critical weaknesses

W1. It remains unclear why better alignment between surrogate gradients and hard-spike forward dynamics would necessarily improve the adversarial robustness of SNNs. Intuitively, if the surrogate gradients more accurately reflect spike flips during forward propagation, attackers could also exploit these more informative gradients to generate stronger adversarial examples. This concern is also supported by the findings of [1], which demonstrate that improved surrogate-gradient approximations can lead to more effective adversarial attacks.

W2. The ablation studies in Figure 8 show that using only the dynamic response leads to really bad performance, while the main results in Table 1 indicate that using only the original response yields poor adversarial robustness. Surprisingly, simply averaging these two responses results in significant performance improvements. The authors should provide a more detailed explanation of why combining two individually ineffective responses leads to such substantial gains.

[1] Towards Reliable Evaluation of Adversarial Robustness for Spiking Neural Networks. CVPR 2026.

Questions for authors

Please clarify the Weaknesses part as well as the following question:

Q1. According to the paper, the importance weight wkw_k is computed using gradients and plays a crucial role in determining β^\hat{\beta} (Eq. (13)). However, computing wkw_k appears to require surrogate gradients, which the authors argue are inconsistent with the actual hard-spike forward dynamics. This raises a concern about whether such potentially inaccurate gradient-based weights can reliably guide the estimation of β^\hat{\beta}.

I would be willing to reconsider my rating if all my concerns are adequately addressed.

10 comments
AuthorsOpenReview ↗

Response to Reviewer 4y1W (part 1/4)

We appreciate reviewer 4y1W’s thoughtful comments and helpful suggestions for strengthening the presentation and evaluation of our work.


Weakness 1: W1. It remains unclear why better alignment between surrogate gradients and hard-spike forward dynamics would necessarily improve the adversarial robustness of SNNs. Intuitively, if the surrogate gradients more accurately reflect spike flips during forward propagation, attackers could also exploit these more informative gradients to generate stronger adversarial examples. This concern is also supported by the findings of [1], which demonstrate that improved surrogate-gradient approximations can lead to more effective adversarial attacks.


Response to Weakness 1:

Thanks for raising this important question.

1. Why better alignment can help robust learning

The key is that surrogate gradients determine which membrane-state changes receive attention during training. For example, a state may contribute little under a clean input, yet undergo a spike flip under a small perturbation that changes the final decision. If the fixed surrogate assigns a weak response to that state, backpropagation can underestimate its importance. The network can then fit clean samples well while giving these vulnerable states too little attention.

Fare uses actual flips and their decision relevance to redirect some backward sensitivity toward these overlooked states, while retaining the original response to support basic learning. This gives parameter updates more useful information about where perturbations affect decisions. Better alignment therefore provides a mechanism for improving learning, while the resulting robustness is checked experimentally. The allocation changes in Figure 6, the feedback-source ablation in Figure 9 (Newly added in the revised manuscript), and the strong-attack results jointly support this explanation.

2. We also give attackers access to Fare’s gradients

We considered this possibility when designing our evaluation, which is why we included “White-box robustness under alternative surrogate-gradient access” in Section 5.1 (p. 7). For each fixed trained model, attackers use Triangle, Sigmoid, Rectangle, and their Fare-adjusted versions to generate adversarial examples. In simple terms, attackers can use the same flip-aware gradients that Fare provides, alongside alternative backward rules.

This directly tests whether changing the attack gradient reveals many previously missed adversarial examples. Across the FGSM and PGD-10 settings in Figure 5, the maximum accuracy reductions are 2.80% and 2.30%, respectively. For example, the CIFAR-10 Fare-Triangle model’s PGD accuracy changes from 33.89% under its native rule to a minimum of 31.82%. There is no abrupt collapse across these tested backward rules, indicating that the model’s performance does not depend on the attacker choosing a particularly ineffective gradient.

3. Stronger attacks provide a further check

The finding in [1] that better gradients can produce stronger attacks is consistent with our reasoning. At attack time, better gradients help expose weaknesses in an already fixed model. During training, Fare uses improved sensitivity information to change the model that is learned. These two effects can coexist.

We further evaluate the final models using longer multi-restart PGD, APGD-CE/DLR, and per-example aggregation across attack losses and backward rules. These additions are highlighted in Section 5.3, “Gradient Obfuscation Checking” (pp. 9–10), with results in Table 3. APGD does uncover weaknesses missed by PGD-10, so we report the aggregated results to reflect this stronger evaluation. Increasing the white-box search from W100W_{100} to W200W_{200} reduces accuracy by a further 0.09-0.50%. Adding SPSA and Square with 10000 queries causes at most another 0.50% reduction.

Together, these results support that Fare’s robustness under the tested attacks does not come from preventing attackers from effectively using its backward gradients.

For convenience, we reproduce the added discussion (Section 5.3) in the next part of our response.

AuthorsOpenReview ↗

Response to Reviewer 4y1W (part 2/4)

Continuation of our response to Weakness 1

5.3 Gradient Obfuscation Checking

Since Fare directly modifies surrogate gradients, we examine whether its robustness depends on difficulties in gradient-based adversarial search. Following the diagnostics of [2], we assess attack strength, backward-rule dependence, and gradient-free search as follows.

Attack strength and backward access. Figure 4 shows that iterative PGD is more effective than FGSM and white-box attacks outperform transfer-based black-box attacks. PGD accuracy decreases as the perturbation budget increases and approaches zero at large budgets. Figure 5 further shows that replacing the attack backward rule causes no abrupt accuracy collapse while the model remains fixed.

Strengthened white-box search. Table 3 includes multi-restart PGD and APGD-CE/DLR. APGD yields lower accuracy than PGD, motivating per-example aggregation of attack successes across losses and backward rules. Expanding the white-box ensemble from W100W_{100} to W200W_{200} reduces accuracy by only 0.09-0.50% across all valid configurations. This limited additional reduction suggests that the aggregated search is approaching saturation under the tested budgets.

Gradient-free verification. Furthermore, as shown in Table 3, adding SPSA and Square Attack to W200W_{200} reduces accuracy by only 0.03-0.50 percentage points. Increasing the Square query budget from 5000 to 10000 yields a further reduction of 0.01-0.30%. These independent searches identify few additional failures after strengthened white-box evaluation. Taken together, the checks show that Fare does not rely on gradient obfuscation for robustness.

Table 3. Strong-attack evaluation of Fare (%).

Attack / EnsembleCIFAR-10CIFAR-100DVSCIFAR-10SeqCIFAR-10
Base surrogateTriSigRecTriSigRecTriSigRecTriSigRec
PGD−10PGD-1033.8929.9124.7514.0712.829.2562.9050.00N/A13.039.097.36
SPSASPSA36.9631.0928.4819.1214.769.9964.4052.10N/A15.9712.478.01
$PGD-10 \cup SPSA $33.6029.8724.7214.0712.589.1662.7049.60N/A13.008.947.18
PGD−50/R10PGD-50 / R1031.1926.4922.8813.2511.108.1660.0049.10N/A11.988.216.10
PGD−100/R10PGD-100 / R1031.0826.4222.7313.1710.988.0459.8049.00N/A11.747.975.99
APGD−CEAPGD-CE19.2414.5812.044.023.763.1023.5021.70N/A3.102.191.47
APGD−DLRAPGD-DLR27.5821.4118.512.541.981.7229.7024.70N/A2.761.270.99
W100W_{100}18.7214.1711.682.311.821.5122.7020.90N/A2.511.120.83
W200W_{200}18.2513.7711.272.171.641.3622.3020.40N/A2.381.030.74
W200∪SPSA∪S5kW_{200} \cup SPSA \cup S_{5k}18.1713.6511.212.121.631.3322.2020.10N/A2.330.990.72
W200∪SPSA∪S10kW_{200} \cup SPSA \cup S_{10k}18.1413.6311.202.091.611.3021.9019.90N/A2.290.970.70

Reference

[2] Athalye, Anish, et al. Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples. ICML, 2018.

AuthorsOpenReview ↗

Response to Reviewer 4y1W (part 3/4)


Weakness 2: W2. The ablation studies in Figure 8 show that using only the dynamic response leads to really bad performance, while the main results in Table 1 indicate that using only the original response yields poor adversarial robustness. Surprisingly, simply averaging these two responses results in significant performance improvements. The authors should provide a more detailed explanation of why combining two individually ineffective responses leads to such substantial gains.


Response to Weakness 2:

Thanks for asking us to explain this.

First, allow us to clarify, each response is useful, rather than “ineffective”. But they have different limitations. The original response supports strong clean accuracy, while providing limited guidance about perturbation-induced flips. The dynamic response already improves robustness on the non-event datasets. For example, using it alone raises SeqCIFAR-10 PGD accuracy from 0.04% to 9.47%. The severe collapse occurs on DVSCIFAR-10. These results suggest that the two responses support different learning needs, which helps explain their benefit when combined.

Their main advantage is that they compensate for weaknesses at different membrane states. The original response consistently supports learning near the firing threshold. The dynamic response extends sensitivity toward important flip regions, but changing its range can weaken the response in regions needed for ordinary classification. A simple analogy is a learning signal passing through several gates. One response keeps some gates open but leaves others nearly closed, while the other response has a different set of weak gates. Combining them can keep the signal flowing through the whole path. This matters because gradients travel across multiple layers and timesteps, where a weak response at one important point can limit learning along an entire path. The combined response can therefore improve gradient propagation substantially, even though it remains a simple average locally.

Furthermore, at the formula level, you are absolutely right that Fare is “simply averaging these two responses.” However, this operation happens at each neuron during training, rather than after two models have finished training. The combined response shapes parameter updates, which change the network’s membrane states and spike behavior. Later calibration then uses this new flip feedback to adjust the dynamic response ranges. In other words, this “averaging” repeats over many training rounds, and earlier “averaging” influences what gets averaged later. This ongoing interaction changes the entire learning process. The final accuracy therefore need not fall between the accuracies obtained using either response alone.

Reviewer_fmMxOpenReview ↗
Rating
2
Confidence
4(+2)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the mismatch between surrogate-gradient sensitivity and actual adversarial responses in spiking neural networks (SNNs). The authors identify a phenomenon termed adversarial sensitivity misallocation, where conventional surrogate gradients fail to adequately capture perturbation-induced, decision-relevant spike flips.

To address this issue, the paper proposes Flip-Aware Reallocation (FARE), which dynamically adjusts surrogate-gradient response ranges using decision-weighted spike-flip statistics. The method combines the adapted response with the original surrogate function while preserving the hard-spike forward dynamics. Experiments across five datasets demonstrate substantial improvements in adversarial robustness with relatively small reductions in clean accuracy.

Critical strengths
  1. Interesting motivation and insightful analysis. The paper investigates an important limitation of surrogate-gradient learning from the perspective of adversarial sensitivity allocation. The analysis of forward-backward damage discrepancies, global scaling effects, and spike-flip distributions provides a coherent motivation for the proposed method.

  2. Simple and potentially generalizable design. FARE modifies the surrogate-gradient response rather than the forward neuronal dynamics or network architecture. Its compatibility with Triangle, Sigmoid, and Rectangle surrogates is encouraging, and the absence of additional inference overhead is a practical advantage.

  3. Strong empirical improvements. The reported robustness gains are substantial across static, sequential, event-based, and large-scale classification tasks. For example, PGD accuracy improves from 0.03% to 33.89% on CIFAR-10 and from 5.00% to 62.90% on DVSCIFAR-10, while largely preserving clean accuracy. The additional analysis of sensitivity distributions and alternative backward rules further supports the proposed approach.

Critical weaknesses
  1. The adversarial robustness evaluation could be stronger. The experiments mainly rely on FGSM and 10-step PGD, which may not be sufficient to rule out gradient masking, especially since FARE directly modifies surrogate gradients. Recent work, such as Towards Reliable Evaluation of Adversarial Robustness for Spiking Neural Networks (Wang et al., CVPR 2026), has shown that conventional surrogate-based attacks may overestimate SNN robustness. Evaluations with longer multi-restart PGD and stronger SNN-specific attacks would make the results more convincing.

  2. More ablations are needed to verify the effectiveness of the proposed mechanism. The current experiments mainly study the quantile parameter and the contribution of the original surrogate function. It remains unclear whether the improvements come specifically from decision-weighted spike-flip feedback or simply from adjusting the surrogate response width. Comparisons with uniform decision weights, membrane-potential-based adaptation, symmetric adaptive widths, and simpler dynamic surrogate functions would help clarify this point.

  3. The comparison with related work is not sufficiently complete. Recent methods such as A²SG (ICML 2026) and MPD-SGR (AAAI 2026) are closely related to adaptive surrogate gradients and robustness, but their differences from FARE are not clearly discussed. Also, Table 1 compares against STOD without orthogonal kernels, while the complete STOD achieves higher FGSM and PGD accuracy on CIFAR-10. Although FARE has advantages in clean accuracy and inference cost, the SOTA claims should be more carefully stated.

  4. The connection between the theoretical analysis and the actual implementation is not fully clear. Equation (11) derives a conditionally optimal surrogate response, but FARE uses weighted quantiles and EMA to approximate the response range. The paper does not sufficiently explain why this approximation should approach the theoretical objective. In addition, the use of globally shared response ranges may not adequately account for differences across layers and timesteps. More analysis of these design choices would strengthen the technical contribution.

  5. The training overhead is relatively high, and the statistical evaluation is limited. According to Table 6, FARE substantially increases training time and approximately doubles or more the peak memory usage in several settings. Although inference cost remains unchanged, this additional training cost should be better discussed. Moreover, results from multiple independent training seeds are missing, making it difficult to assess the stability of the reported improvements.

Questions for authors

See weaknesses.

10 comments
AuthorsOpenReview ↗

Response to Reviewer fmMx (part 1/5)

We thank reviewer fmMx for the careful reading and constructive feedback, which have helped us improve the paper.


Weakness 1: The adversarial robustness evaluation could be stronger. The experiments mainly rely on FGSM and 10-step PGD, which may not be sufficient to rule out gradient masking, especially since FARE directly modifies surrogate gradients. Recent work, such as Towards Reliable Evaluation of Adversarial Robustness for Spiking Neural Networks (Wang et al., CVPR 2026), has shown that conventional surrogate-based attacks may overestimate SNN robustness. Evaluations with longer multi-restart PGD and stronger SNN-specific attacks would make the results more convincing.


Response to Weakness 1:

Thanks for emphasizing this point.

We have expanded the evaluation to check whether Fare’s robustness survives stronger and more varied attack settings. The additional tests include 50- and 100-step PGD with 10 random restarts, APGD-CE/DLR, alternative surrogate backward rules, and gradient-free SPSA and Square attacks. We also combine attack successes per example, so an input counts as robust only if it survives every included attack.

These checks are presented in newly added section: Section 5.3, “Gradient Obfuscation Checking” (pp. 9–10), which is highlighted in our revised manuscript. Together, they provide substantially stronger evidence than FGSM and PGD-10 alone. For convenience, we reproduce the added discussion below:

5.3 Gradient Obfuscation Checking

Since Fare directly modifies surrogate gradients, we examine whether its robustness depends on difficulties in gradient-based adversarial search. Following the diagnostics of [1], we assess attack strength, backward-rule dependence, and gradient-free search as follows.

Attack strength and backward access. Figure 4 shows that iterative PGD is more effective than FGSM and white-box attacks outperform transfer-based black-box attacks. PGD accuracy decreases as the perturbation budget increases and approaches zero at large budgets. Figure 5 further shows that replacing the attack backward rule causes no abrupt accuracy collapse while the model remains fixed.

Strengthened white-box search. Table 3 includes multi-restart PGD and APGD-CE/DLR. APGD yields lower accuracy than PGD, motivating per-example aggregation of attack successes across losses and backward rules. Expanding the white-box ensemble from W100W_{100} to W200W_{200} reduces accuracy by only 0.09-0.50% across all valid configurations. This limited additional reduction suggests that the aggregated search is approaching saturation under the tested budgets.

Gradient-free verification. Furthermore, as shown in Table 3, adding SPSA and Square Attack to W200W_{200} reduces accuracy by only 0.03-0.50 percentage points. Increasing the Square query budget from 5000 to 10000 yields a further reduction of 0.01-0.30%. These independent searches identify few additional failures after strengthened white-box evaluation. Taken together, the checks show that Fare does not rely on gradient obfuscation for robustness.

Table 3. Strong-attack evaluation of Fare (%).

Attack / EnsembleCIFAR-10CIFAR-100DVSCIFAR-10SeqCIFAR-10
Base surrogateTriSigRecTriSigRecTriSigRecTriSigRec
PGD−10PGD-1033.8929.9124.7514.0712.829.2562.9050.00N/A13.039.097.36
SPSASPSA36.9631.0928.4819.1214.769.9964.4052.10N/A15.9712.478.01
$PGD-10 \cup SPSA $33.6029.8724.7214.0712.589.1662.7049.60N/A13.008.947.18
PGD−50/R10PGD-50 / R1031.1926.4922.8813.2511.108.1660.0049.10N/A11.988.216.10
PGD−100/R10PGD-100 / R1031.0826.4222.7313.1710.988.0459.8049.00N/A11.747.975.99
APGD−CEAPGD-CE19.2414.5812.044.023.763.1023.5021.70N/A3.102.191.47
APGD−DLRAPGD-DLR27.5821.4118.512.541.981.7229.7024.70N/A2.761.270.99
W100W_{100}18.7214.1711.682.311.821.5122.7020.90N/A2.511.120.83
W200W_{200}18.2513.7711.272.171.641.3622.3020.40N/A2.381.030.74
W200∪SPSA∪S5kW_{200} \cup SPSA \cup S_{5k}18.1713.6511.212.121.631.3322.2020.10N/A2.330.990.72
W200∪SPSA∪S10kW_{200} \cup SPSA \cup S_{10k}18.1413.6311.202.091.611.3021.9019.90N/A2.290.970.70

Reference

[1] Athalye, Anish, et al. Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples. ICML, 2018.

AuthorsOpenReview ↗

Response to Reviewer fmMx (part 2/5)


Weakness 2: More ablations are needed to verify the effectiveness of the proposed mechanism. The current experiments mainly study the quantile parameter and the contribution of the original surrogate function. It remains unclear whether the improvements come specifically from decision-weighted spike-flip feedback or simply from adjusting the surrogate response width. Comparisons with uniform decision weights, membrane-potential-based adaptation, symmetric adaptive widths, and simpler dynamic surrogate functions would help clarify this point.


Response to Weakness 2:

Thanks for asking us to clarify what the flip feedback and decision weighting contribute beyond adaptive response widths.

We have added and highlighted a feedback-source ablation on CIFAR-10 in the revised manuscript. All four configurations dynamically adjust their response widths using the same update and fusion rules. We change the information used to estimate those widths:

  • A: All sampled membrane-to-threshold distances, with uniform weights.
  • B: The same membrane states as A, with decision weights.
  • C: Only distances associated with actual spike flips, with uniform weights.
  • Fare: The same actual flips as C, with decision weights.

This gives us two direct comparisons.

  • 1: A versus C tests whether selecting actual flips helps

  • 2: C versus Fare tests whether decision weighting adds useful information.

The added analysis appears in Section 5.2, “Role of weighted flip feedback” (p. 9, Figure 9). We reproduce the relevant manuscript text below:

Role of weighted flip feedback

On CIFAR-10, we compare four response-range estimators with identical quantile updates and unit-mass surrogate fusion. The control group comprised of three settings of A, B, and C: A uses unweighted membrane-to-threshold distances, B applies decision weights to the same membrane statistics, C uses unweighted spike-flip distances, and Fare uses decision-weighted spike-flip distances. Details of the controls are provided in Appendix E.3. All four configurations dynamically adapt their widths, isolating the contributions of flip selection and decision weighting. Figure 9 reports accuracy and mean dTVd_{\mathrm{TV}}, with the latter averaged over FGSM and PGD. Compared with A, C increases PGD accuracy from 9.01% to 29.44% and reduces mean dTVd_{\mathrm{TV}} from 0.320 to 0.258, supporting actual flips as informative feedback for identifying response regions. B also improves robustness, but remains below C, and both reduce clean accuracy relative to A. Adding decision weighting to C further improves clean, FGSM, and PGD accuracy by 3.08%, 3.52%, and 4.45%, respectively, while lowering mean dTVd_{\mathrm{TV}} to 0.235. These results support complementary roles for flip selection and decision weighting beyond width adaptation alone. In particular, weighting actual flips by their decision relevance plays an important role in preserving clean accuracy while further improving robustness and sensitivity reallocation effectiveness.

Source data of Figure 9 follow belows:

ConfigurationClean (%)FGSM (%)PGD (%)Mean dTVd_{\mathrm{TV}}
A92.5123.919.010.320
B91.0441.4820.750.299
C90.4655.1829.440.258
Fare93.5458.7033.890.235
AuthorsOpenReview ↗

Response to Reviewer fmMx (part 3/5)


Weakness 3: The comparison with related work is not sufficiently complete. Recent methods such as A²SG (ICML 2026) and MPD-SGR (AAAI 2026) are closely related to adaptive surrogate gradients and robustness, but their differences from FARE are not clearly discussed. Also, Table 1 compares against STOD without orthogonal kernels, while the complete STOD achieves higher FGSM and PGD accuracy on CIFAR-10. Although FARE has advantages in clean accuracy and inference cost, the SOTA claims should be more carefully stated.


Response to Weakness 3:

Thanks for pointing out the deficiencies of this work toward existing SNN work searching. We have added A²SG [1] and MPD-SGR [2] to our Related Work Section, located at Section 6, p. 10, revised content has been highlighted. Besides, we appreciate the opportunity to clarify the differences between Fare and the methods you mentioned.

1. A²SG and Fare

A²SG adapts surrogate windows and asymmetric profiles to reduce spatial gradient variation. Its focus is training stability and generalization, and its robustness experiments consider internal spike-deletion noise. Fare addresses input-space adversarial perturbations and uses a different source of feedback: actual spike flips between clean and perturbed trajectories, weighted by their decision relevance. The key difference is therefore what guides adaptation, beyond the shared idea of adjusting surrogate windows.

2. MPD-SGR and Fare

MPD-SGR adds a regularization loss that changes the membrane-potential distribution’s overlap with a given surrogate response region. The concept of MPD-SGR is quite like TGO [3], which is already cited and compared in our original manuscript (Table 1, pp. 6). Fare adjusts the surrogate response itself using observed adversarial flips. Put simply, MPD-SGR changes where membrane states fall relative to the response window, while Fare adjusts the window using flip feedback. Fare also retains the original response as a threshold-local learning anchor and preserves total surrogate mass.

3. STOD Comparison

For STOD, the original paper [4] explicitly states that retaining the orthogonal kernels introduces additional inference overhead. Directly comparing that version with Fare would therefore involve models with different inference-time complexity, which we consider unfair. Importantly, the STOD authors themselves use STOD w.o. OK (STOD without Orthogonal Kernel), with the orthogonal kernels removed at inference, in their main comparison against SOTA methods. Our choice follows the comparison setting adopted in the original paper. We therefore consider using this variant fully justified and appropriate.

References

[1] Kang, Yechan, et al. A2SG: Adaptive and Asymmetric Surrogate Gradients for Training Deep Spiking Neural Networks. ICML, 2026.

[2] Jiang, Runhao, et al. Mpd-sgr: Robust spiking neural networks with membrane potential distribution-driven surrogate gradient regularization. AAAI, 2026.

[3] Wang, Shuai, et al. Robust spiking neural networks against adversarial attacks. ICLR, 2026.

[4] Zhang, Desong, et al. Breaking Gradient Temporal Collinearity for Robust Spiking Neural Networks. ICLR, 2026.

Reviewer_SU27OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The submission investigates the vulnerability of surrogate-gradient (SG) trained Spiking Neural Networks to adversarial attacks. The authors argue that conventional surrogate functions (triangular, sigmoid, etc.) have fixed activation widths centered at the firing threshold, meaning backward sensitivity fails to cover membrane potentials where spike flips actually happen during an adversarial attack.

To tackle this, they first show empirically that a simple scalar multiplier cannot fix this gradient misallocation. They then propose FARE, which periodically gathers spike flips from perturbed calibration samples to estimate adaptive surrogate ranges (β+\beta_+, β−\beta_-) via quantiles, combining this dynamic shape with a standard static surrogate. Empirical results are reported on CIFAR-10/100, DVS-CIFAR10, and ImageNet against FGSM and 10-step PGD.

Critical strengths

(1) The diagnostic experiments in Section 3 are convincing. Testing across 108 experimental setups and showing that even an optimal scalar multiplier leaves high residual errors (up to ~0.70) provides solid empirical support for their claim. The total variation distance (dTVd_{TV}) between the surrogate mass and the actual flip distribution clearly illustrates that the issue is spatial misallocation rather than gradient scale.

(2) Maintaining clean accuracy while gaining adversarial robustness is notoriously hard in SNNs, where standard adversarial training often costs 5-15% in clean performance. FARE seems to circumvent this tradeoff on CIFAR-10 (retaining ~93.7% clean accuracy while improving PGD-10). Scaling this up to ImageNet on Spiking NF-ResNet18 is also a welcome validation that is often missing in related literature.

Overall, the paper is well-written and easy to follow.

Critical weaknesses

(1) The primary concern lies in the evaluation protocol. Relying almost exclusively on FGSM and 10-step PGD (α=0.01\alpha=0.01) is insufficient for claiming adversarial robustness in SNNs. The SNN community has repeatedly seen defenses fail due to obfuscated or vanishing surrogate gradients (gradient masking). Without evaluation under stronger attack benchmarks—specifically AutoAttack / APGD, multi-restart PGD with 50-100 steps, or gradient-free checks like SPSA—it is impossible to verify whether FARE genuinely expands the classification margin or merely degrades the quality of short-step first-order gradients.

(2) Some calibration details are missing. Appendix C mentions flips are collected "every K mini-batches" but the actual value of K isn't stated anywhere. Also, Table 6 shows training time increases by up to 235% with doubled GPU memory. Since FARE actively perturbs samples during training, this is basically lightweight adversarial training—comparing it to standard clean training without matched compute isn't really fair.

Questions for authors

(1) What is the exact value of K used in Tables 1 and 2?

(2) Why does a small calibration budget (ϵcal=2/255\epsilon_{cal}=2/255) work when the test-time attack uses 8/2558/255 or larger? I'd expect the estimated ranges to shift substantially.

(3) It would really help to see CIFAR-10 numbers under AutoAttack or APGD to rule out gradient masking.

(4) In Fig. 8(c), the "Dynamic SG only" setup collapses on DVSCIFAR-10 (clean ~38%, PGD ~0.1%). Any intuition for why the flip-aware branch is so unstable on event data without the static anchor?

5 comments
AuthorsOpenReview ↗

Response to Reviewer SU27 (part 1/4)

We sincerely thank reviewer SU27 for the positive assessment and recognition of our contribution, alongside the constructive suggestions that have helped us further strengthen the paper.


Weakness 1: The primary concern lies in the evaluation protocol. Relying almost exclusively on FGSM and 10-step PGD (α=0.01\alpha = 0.01) is insufficient for claiming adversarial robustness in SNNs. The SNN community has repeatedly seen defenses fail due to obfuscated or vanishing surrogate gradients (gradient masking). Without evaluation under stronger attack benchmarks—specifically AutoAttack / APGD, multi-restart PGD with 50-100 steps, or gradient-free checks like SPSA—it is impossible to verify whether FARE genuinely expands the classification margin or merely degrades the quality of short-step first-order gradients.

Question 3: It would really help to see CIFAR-10 numbers under AutoAttack or APGD to rule out gradient masking.


Response to Weakness 1 & Question 3:

Thank you for raising this concern about gradient masking (i.e., gradient obfuscation). We recognize that checking for gradient obfuscation is essential for reliable robustness evaluation in SNNs. We address this issue by adding Section 5.3, “Gradient Obfuscation Checking” (pp. 9-10), which is highlighted in the revised manuscript. In this subsection, following the diagnostic checks described by [1], we evaluate Fare using additional attacks, including longer PGD runs, multiple random restarts, AutoPGD, and gradient-free attacks. Together, these analyses and results provide strong evidence that Fare’s robustness does not rely on gradient obfuscation.

For convenience, we reproduce the section added to the revised manuscript below.

5.3 Gradient Obfuscation Checking

Since Fare directly modifies surrogate gradients, we examine whether its robustness depends on difficulties in gradient-based adversarial search. Following the diagnostics of [1], we assess attack strength, backward-rule dependence, and gradient-free search as follows.

Attack strength and backward access. Figure 4 shows that iterative PGD is more effective than FGSM and white-box attacks outperform transfer-based black-box attacks. PGD accuracy decreases as the perturbation budget increases and approaches zero at large budgets. Figure 5 further shows that replacing the attack backward rule causes no abrupt accuracy collapse while the model remains fixed.

Strengthened white-box search. Table 3 includes multi-restart PGD and APGD-CE/DLR. APGD yields lower accuracy than PGD, motivating per-example aggregation of attack successes across losses and backward rules. Expanding the white-box ensemble from W100W_{100} to W200W_{200} reduces accuracy by only 0.09-0.50% across all valid configurations. This limited additional reduction suggests that the aggregated search is approaching saturation under the tested budgets.

Gradient-free verification. Furthermore, as shown in Table 3, adding SPSA and Square Attack to W200W_{200} reduces accuracy by only 0.03-0.50%. Increasing the Square query budget from 5000 to 10000 yields a further reduction of 0.01-0.30%. These independent searches identify few additional failures after strengthened white-box evaluation. Taken together, the checks show that Fare does not rely on gradient obfuscation for robustness.

Table 3. Strong-attack evaluation of Fare (%).

Attack / EnsembleCIFAR-10CIFAR-100DVSCIFAR-10SeqCIFAR-10
Base surrogateTriSigRecTriSigRecTriSigRecTriSigRec
PGD−10PGD-1033.8929.9124.7514.0712.829.2562.9050.00N/A13.039.097.36
SPSASPSA36.9631.0928.4819.1214.769.9964.4052.10N/A15.9712.478.01
$PGD-10 \cup SPSA $33.6029.8724.7214.0712.589.1662.7049.60N/A13.008.947.18
PGD−50/R10PGD-50 / R1031.1926.4922.8813.2511.108.1660.0049.10N/A11.988.216.10
PGD−100/R10PGD-100 / R1031.0826.4222.7313.1710.988.0459.8049.00N/A11.747.975.99
APGD−CEAPGD-CE19.2414.5812.044.023.763.1023.5021.70N/A3.102.191.47
APGD−DLRAPGD-DLR27.5821.4118.512.541.981.7229.7024.70N/A2.761.270.99
W100W_{100}18.7214.1711.682.311.821.5122.7020.90N/A2.511.120.83
W200W_{200}18.2513.7711.272.171.641.3622.3020.40N/A2.381.030.74
W200∪SPSA∪S5kW_{200} \cup SPSA \cup S_{5k}18.1713.6511.212.121.631.3322.2020.10N/A2.330.990.72
W200∪SPSA∪S10kW_{200} \cup SPSA \cup S_{10k}18.1413.6311.202.091.611.3021.9019.90N/A2.290.970.70

Reference

[1] Athalye, Anish, et al. Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples. ICML, 2018.

AuthorsOpenReview ↗

Response to Reviewer SU27 (part 2/4)


Weakness 2: Some calibration details are missing. Appendix C mentions flips are collected "every K mini-batches" but the actual value of K isn't stated anywhere. Also, Table 6 shows training time increases by up to 235% with doubled GPU memory. Since FARE actively perturbs samples during training, this is basically lightweight adversarial training—comparing it to standard clean training without matched compute isn't really fair.

Question 1: What is the exact value of K used in Tables 1 and 2?


Response to Weakness 2 & Question 1:

Thanks for your commments. We address them separately as follows:

1. Calibration Interval

We sincerely apologize for this minor mistake, and thank you very much for pointing this out. Fare collects spike-flip statistics every Kcal=10K_{\mathrm{cal}}=10 mini-batches for all experiments in Tables 1 and 2. We omitted this numerical value from Appendix C and have now corrected it. We also renamed the interval from KK to KcalK_{\mathrm{cal}} to distinguish it from the number of attack iterations. This point has been corrected and highlighted in the revised manuscript (p. 15). It now states:

During training, Fare collects spike-flip statistics every KcalK_{\mathrm{cal}} mini-batches, with Kcal=10K_{\mathrm{cal}}=10 for all experiments.

2. Training Time and Memory

The additional cost in Table 6 comes from periodic flip calibration, including perturbation generation, comparison of clean and perturbed spike trajectories, and estimation of decision relevance. These operations require additional temporal computation and intermediate-state storage, increasing training time and peak GPU memory. The maximum time increase of approximately 235% occurs on DVSCIFAR-10, and the overhead varies across datasets.

All of this additional work is confined to training. Once training finishes, the calibration branch is inactive, and Fare retains the original hard-spike inference graph without additional model parameters, inference operations, or latency. This makes the design particularly useful when a trained model will perform inference repeatedly.

The additional training cost also accompanies substantial robustness gains. Under the PGD-10 evaluation reported in the original manuscript, Fare improves DVSCIFAR-10 accuracy from 5.00% to 62.90% while maintaining the same 76.30% clean accuracy. Fare therefore offers a practical trade-off: higher one-time training cost for improved robustness, just like many other works did, with unchanged deployment efficiency.

3. Comparison with Adversarial Training

Fare uses perturbed samples as diagnostic cases: they reveal where spike flips occur and help determine which flips matter to the decision. Fare then uses this feedback to adjust the surrogate response ranges. The parameter update still uses the supervised loss on clean inputs. Standard adversarial training directly includes a classification loss on adversarial examples in that update. A simple analogy is using a stress test to identify where a structure needs reinforcement. The test guides the reinforcement design, and collecting that feedback is part of the method itself. This also matters when choosing comparisons.

Suppose a paper reports both method X and X+AT. Adding a separate AT procedure to X creates an additional defense configuration with a different objective and training process. Automatically requiring this extra procedure for selected competitors can make the standalone-method comparison unfair to Fare.

Furthermore, our comparison already includes methods whose training uses adversarial examples. In particular, the SR results in Table 1 correspond to its adversarially trained SR* configuration. On CIFAR-10, Fare achieves 93.54% clean accuracy and 33.89% PGD accuracy, compared with 85.91% and 30.54% for SR. Thus, Fare’s advantage also holds against this adversarially trained comparator.

AuthorsOpenReview ↗

Response to Reviewer SU27 (part 3/4)


Question 2: Why does a small calibration budget (2/255) work when the test-time attack uses 8/255 or larger? I'd expect the estimated ranges to shift substantially.


Response to Question 2:

Thanks for reviewer’s insightful comment.

The calibration budget and the test-time attack budget serve different purposes. Calibration provides feedback for adjusting the learning rule, while the test budget determines how strongly the trained model is attacked. These budgets therefore do not need to match. Previous SNN studies also report benefits from training with weaker perturbations and testing with stronger ones. For example, SR [1] uses PGD perturbations at 2/2552/255 during training and evaluates robustness at 8/2558/255. HoSNN [2] similarly reports robustness gains against 8/2558/255 attacks after training with FGSM perturbations at 2/2552/255.

For Fare, 2/2552/255 is an empirical default for collecting flip feedback. A larger calibration budget can reach different membrane states and change the estimated ranges. The goal is to obtain useful feedback for learning a robust model, so the calibration budget does not set an upper limit on the test-time perturbations it can withstand. Figure 4 in our manuscript already shows that Fare trained with this default improves robustness across multiple test budgets.

To examine this choice directly, we ran an additional calibration-budget ablation on CIFAR-10 using VGG11 and Triangle. All FGSM and PGD evaluations below use a fixed test budget of 8/2558/255.

Calibration budgetClean (%)FGSM at 8/255 (%)PGD at 8/255 (%)
2/255 (default Fare)93.5458.7033.89
4/25589.2964.8837.01
8/25581.1268.2039.19
16/25578.6662.1836.64

Increasing the calibration budget from 2/2552/255 to 8/2558/255 improves adversarial accuracy, but comes with a substantial clean-accuracy cost. The 8/2558/255 setting achieves the highest FGSM and PGD accuracies among the tested budgets, yet its 5.30% PGD gain over the default accompanies a 12.42% drop in clean accuracy. Increasing the budget further to 16/25516/255 reduces both clean and adversarial accuracy. These results support 2/2552/255 as a balanced default, providing substantial robustness gains while largely preserving clean accuracy.

References

[1] Liu, Yujia, et al. Enhancing adversarial robustness in SNNs with sparse gradients. ICML, 2024.

[2] Geng, Hejia, et al. Hosnn: Adversarially-robust homeostatic spiking neural networks with adaptive firing thresholds. TMLR, 2025.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overall Response

We sincerely thank all reviewers for their thoughtful feedback. We have strengthened the manuscript and provided additional results addressing the main concerns:

  • Stronger robustness evaluation : addressing reviewer SU27’s Weakness 1 and Question 3, reviewer fmMx’s Weakness 1, and reviewer 4y1W’s Weakness 1. The revised manuscript includes an expanded evaluation supporting Fare’s robustness beyond the original short-step attacks and checking its dependence on gradient obfuscation.

  • Direct evidence for the proposed mechanism: addressing reviewer fmMx’s Weakness 2 and reviewer 4y1W’s Question 1. A new manuscript ablation supports the contributions of actual spike-flip feedback and decision weighting beyond response-width adaptation alone.

  • Additional stability and reproducibility evidence: addressing reviewer fmMx’s Weakness 5 and reviewer SU27’s Weakness 2 and Questions 1-2. Our responses provide independent-seed results and a calibration-budget study, alongside clearer calibration settings and training-cost explanations.

The new robustness checks and mechanism ablation, together with the corresponding appendix updates, are highlighted in the revised manuscript. The additional seed and calibration-budget results are provided in our individual responses. Together, these updates strengthen the experimental support for Fare and clarify its practical trade-offs.

Best regards

Submission 3060 Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.