acceptodds
Under review as a conference paper at ICLR 2027

Dormant in ANNs, Activated in SNNs: Stealthy Backdoor Attacks via ANN-to-SNN Conversion

Abstract

Spiking neural networks (SNNs) provide energy-efficient inference through sparse, event-driven computation on neuromorphic hardware. ANN-to-SNN conversion is therefore a widely used route to deploy pretrained artificial neural networks (ANNs) as SNNs without the difficulty of direct SNN training. Existing conversion studies primarily optimize accuracy, latency, and energy, leaving the security of the conversion process itself largely unexplored. We show that this overlooked boundary can activate a backdoor: a released ANN may remain benign on clean and trigger-bearing inputs, while the SNN obtained from the same ANN produces an attacker-specified target prediction. We term this threat A2S-Attack and formulate it as a bilevel optimization problem: the upper level optimizes the released ANN, while the lower level simulates ANN-to-SNN conversion with surrogate calibration data. The resulting objective preserves clean ANN/SNN utility, suppresses the target response in the ANN, and activates it in the SNN under the same trigger. We then evaluate the threat under matched and unknown downstream conversion configurations, private calibration resampling, and ANN-side backdoor screening using a representative training-free conversion pipeline. Across four image datasets, four architectures, and four conversion configurations, matched attack and deployment settings yield ANN targeted attack success rates of at most but SNN attack success rates of –. A single frozen ANN also achieves – SNN attack success across the evaluated downstream configurations, demonstrating that security evaluation must include the converted SNN rather than the released ANN alone.

Primary area:
applications to neuroscience & cognitive science
Keywords:
AI safety, ANN-to-SNN conversion, backdoor attack

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_mtCoOpenReview ↗
Rating
2
Confidence
3
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies a backdoor attack against ANN-to-SNN conversion. The main idea is to embed a backdoor that remains inactive in the original ANN but becomes effective after the model is converted into an SNN. The authors propose A2S-Attack, which uses bilevel optimization to maintain clean accuracy in both models, suppress the attack in the ANN, and activate it in the converted SNN.

The method is evaluated under configuration-aware and configuration-unknown settings using MTH conversion. Experiments on four datasets and four architectures show high attack success rates in the converted SNNs, while the original ANNs remain largely unaffected by the triggers. The results suggest that inspecting the ANN alone may not be sufficient to identify security risks introduced during conversion.

Critical strengths
  1. The threat model is interesting and worth investigating. Unlike conventional backdoor attacks that remain effective across model transformations, this attack is designed to become active only after ANN-to-SNN conversion.

  2. The attack has strong results. The converted SNNs achieve nearly 100% attack success rates in several settings, while the corresponding ANNs show much lower attack success rates. Clean accuracy is also largely preserved, suggesting that the attack can achieve its objective without significantly affecting normal model performance.

  3. The experiments cover a reasonable range of datasets and architectures, including both CNNs and Vision Transformers. The authors also consider different conversion configurations and private calibration data.

Critical weaknesses
  1. The generalization of the attack is still unclear. All experiments are based on MTH conversion, with different choices of timesteps and threshold counts. Since the paper presents this as a security concern for ANN-to-SNN conversion more broadly, I would expect at least some evaluation using other conversion methods, such as rate-based or QCFS-based conversion. Also, in the configuration-unknown setting, the attacker still knows the set of possible configurations and explicitly optimizes over them. This makes the setting less general than it initially appears.

  2. Some important baseline comparisons are missing. The results clearly show that the backdoor behaves differently before and after conversion, but it would be useful to see how benign models behave under the same triggers and conversion settings. Without this comparison, it is difficult to tell how much of the observed behavior comes from the attack itself and how much may be related to conversion-induced errors. A comparison with conventional backdoor attacks using the same converter would also help establish the advantage of the proposed method.

  3. I have some concerns about the stealthiness of the triggers. The CNN experiments use a small patch, while the ViT experiments rely on Gaussian perturbations applied to the entire image. These are quite different types of triggers, and the latter may be more noticeable or affect normal predictions more strongly. The paper does not provide quantitative results on perturbation magnitude or perceptual similarity. In addition, the low robust accuracy of some benign models under triggered inputs suggests that the perturbations themselves can have a considerable impact. More analysis would be helpful to support the claims about stealthiness.

  4. The underlying mechanism is not sufficiently explained. Although the method is formulated as bilevel optimization, the actual implementation uses an approximate upper-level gradient that ignores how the calibrated neuron parameters depend on the ANN weights. There is little discussion of how this approximation affects the optimization. More importantly, it remains unclear which part of the conversion process actually causes the dormant backdoor to become active. Ablations on the individual loss terms and analyses of the representations before and after conversion would help explain why the attack works, beyond simply showing that it is effective.

  5. The statistical and defensive evaluations could be stronger. The attack appears to be optimized only once for each configuration, while the repeated experiments mainly vary the calibration samples. This does not tell us much about how stable the attack is across different optimization runs. Also, the detection experiments focus on methods applied to the original ANN. It would be interesting to examine whether the backdoor could be detected by inspecting the converted SNN directly, since this is where the malicious behavior actually appears.

Questions for authors

The author should address the weakness carefully.

9 comments

Reviewer mtCo — W1: Generalization Across Conversion Methods and Configuration Assumptions

1. Generalization Across Conversion Methods

The generalization of the attack is still unclear. All experiments are based on MTH conversion, with different choices of timesteps and threshold counts. Since the paper presents this as a security concern for ANN-to-SNN conversion more broadly, I would expect at least some evaluation using other conversion methods, such as rate-based or QCFS-based conversion.

We thank the reviewer for this valuable suggestion. We agree that evaluating alternative ANN-to-SNN conversion methods is important for assessing the generalizability of A2S-Attack beyond MTH conversion.

To address this concern, we conducted additional experiments using a training-free, QCFS-inspired rate-based conversion scheme with Integrate-and-Fire (IF) neurons, which employs a different neuronal encoding mechanism from MTH without requiring ANN retraining or fine-tuning. The results are reported in Table R4 of our response to Reviewer uBwr (Q3: Generalizability and Mechanism Analysis).

Under Setting I, the A2S-optimized ANNs exhibit negligible targeted attack success rates (ASR) of 0.11% for VGG-16 and 0.01% for ViT-Small, while their IF-converted SNNs achieve 99.82% and 100.00% ASR, respectively, at T=8T=8. Under Setting II, a single ANN per architecture is selected or conditionally refined for robustness across T∈{6,8}T\in\{6,8\}, achieving 98.92%–99.30% ASR for VGG-16 and 99.99% ASR for ViT-Small.

These results demonstrate that conversion-induced backdoor activation is not specific to MTH conversion and can also occur under a distinct rate-based IF conversion scheme. Nevertheless, we do not claim universal effectiveness across all ANN-to-SNN conversion methods.

2. Configuration-Knowledge Assumptions in Setting II

Also, in the configuration-unknown setting, the attacker still knows the set of possible configurations and explicitly optimizes over them. This makes the setting less general than it initially appears.

We acknowledge that Setting II assumes knowledge of the supported configuration set C\mathcal C, while the defender's eventual deployment choice cdeploy∈Cc_{\mathrm{deploy}}\in\mathcal C remains unknown. Thus, Setting II evaluates robustness to an unknown deployment choice within a known candidate set, rather than generalization to arbitrary unseen configurations.

We argue that this assumption is practically meaningful because ANN-to-SNN conversion configurations are constrained by accuracy–efficiency trade-offs. As shown in Table R10 of our response to Reviewer 6MgT (Q2), increasing the MTH threshold count from L=6L=6 to L=10L=10 at T=4T=4 improves VGG-16 accuracy only from 91.32% to 91.87%, while the estimated total energy proxy increases from 71.823 to 138.386 μJ per image. Such trade-offs motivate a limited set of practical configurations. Furthermore, public ANN-to-SNN conversion frameworks, such as SpikingJelly [6, 7], provide documented conversion recipes and reproducible configuration settings, illustrating how candidate configurations can become accessible to downstream users and potential attackers.

Importantly, the attacker releases a single frozen ANN before the defender independently selects its deployment configuration, without the ability to adapt the model afterward. In the main MTH experiments, this single released checkpoint is selected from Setting I candidates in 15/16 cases; only the remaining case receives 100 joint-refinement updates with Algorithm 2. Therefore, Setting II primarily demonstrates cross-configuration transfer, with joint optimization supported as a conditional remedy rather than a uniformly applied training stage. We acknowledge that attacks without restrictions on the downstream configuration space fall outside our current threat model and represent an important direction for future work.

We will revise the manuscript before the rebuttal deadline to incorporate the alternative-conversion results, clarify the configuration-knowledge assumptions, and explicitly state the scope and limitations of Setting II.

Reviewer mtCo – W2: Missing Baseline Comparisons

Reviewer Comment:

Some important baseline comparisons are missing. The results clearly show that the backdoor behaves differently before and after conversion, but it would be useful to see how benign models behave under the same triggers and conversion settings. Without this comparison, it is difficult to tell how much of the observed behavior comes from the attack itself and how much may be related to conversion-induced errors. A comparison with conventional backdoor attacks using the same converter would also help establish the advantage of the proposed method.

Response:

We thank the reviewer for this constructive suggestion. We agree that additional baseline comparisons are important for distinguishing conversion-activated backdoors from ordinary conversion-induced errors and conventional backdoor attacks. To address this concern, we conducted two complementary evaluations.

(1) Comparison with Benign SNNs. We additionally evaluated benign ANNs and their converted SNNs using the same trigger patterns and MTH conversion configurations as A2S-Attack. The results are reported in Table R7 of our response to Reviewer uBwr.

The converted benign SNNs maintain competitive clean accuracy while exhibiting low targeted ASRs (0.01%–4.66%), in contrast to the near-100% ASRs of the corresponding A2S-attacked SNNs. These results indicate that the observed backdoor activation cannot be explained solely by ordinary conversion errors or the presence of the trigger pattern.

(2) Comparison with Conventional Backdoor Attacks. We further evaluated the conventional BadNets attack using the same MTH converter, trigger pattern, and target class as A2S-Attack. To compare detectability during pre-conversion security screening, we also applied Neural Cleanse (NC) to the released ANNs. The results are summarized in Table R19.

Table R19: Comparison with Conventional Backdoor Attacks under MTH Conversion.

CIFAR-10 / ResNet-18, using the same 5×5 trigger, target class, and MTH configuration (T=6,L=8T=6,L=8). A2S results correspond to Setting I. NC is applied to the released ANNs before conversion, with a target-class anomaly-index threshold of τ=2.0\tau=2.0. ACC and ASR are reported in %.

Attack MethodANN ACC ↑ANN ASR ↓SNN ACC ↑SNN ASR ↑NC Target-Class Anomaly IndexNC Detection
BadNets91.90100.0091.739.282.3031Detected
A2S-Attack91.460.0991.2699.920.00Not Detected

As shown in Table R19, BadNets exhibits an active backdoor in the released ANN and is detected by NC, while its targeted ASR decreases to 9.28% after conversion. In contrast, A2S-Attack maintains near-zero ANN ASR (0.09%) and avoids target-class detection by NC, while achieving 99.92% ASR after conversion. This comparison highlights the distinctive advantage of A2S in the evaluated model supply-chain scenario: concealing the malicious target behavior before conversion while activating it in the converted SNN.

We will revise the manuscript before the rebuttal deadline to incorporate these additional baseline comparisons and clarify the distinction between conventional backdoors and conversion-activated attacks.

Reviewer mtCo — W3: Trigger Stealthiness and Perturbation Impact

I have some concerns about the stealthiness of the triggers. The CNN experiments use a small patch, while the ViT experiments rely on Gaussian perturbations applied to the entire image. These are quite different types of triggers, and the latter may be more noticeable or affect normal predictions more strongly. The paper does not provide quantitative results on perturbation magnitude or perceptual similarity.

In addition, the low robust accuracy of some benign models under triggered inputs suggests that the perturbations themselves can have a considerable impact. More analysis would be helpful to support the claims about stealthiness.

We thank the reviewer for this important observation. We first clarify that the stealthiness of A2S-Attack primarily refers to concealing the attacker-specified backdoor behavior in the released ANN, rather than guaranteeing visual imperceptibility of the trigger. Nevertheless, we agree that the perceptual similarity and prediction impact of different trigger types deserve further examination.

(1) Trigger Perturbation and Perceptual Similarity. We acknowledge that the Gaussian trigger implementation was not sufficiently described in Appendix C. Specifically, for ViT-Small/16, the original input xpixelx_{\mathrm{pixel}} is represented in 0–255 pixel-value units, and Gaussian noise δ∼N(0,1)\delta\sim\mathcal{N}(0,1) is added as xpixel+2δx_{\mathrm{pixel}}+2\delta on that 0–255 scale, before subsequent input preprocessing and dataset-specific normalization. Thus, αtrig=2\alpha_{\mathrm{trig}}=2 represents a perturbation standard deviation of 2 pixel-value units, not 2 units on a [0,1][0,1] or standardized input scale. We will clarify this scaling factor and preprocessing order in the revised manuscript.

Although Gaussian triggers cover the entire image, spatial coverage alone does not determine perceptual impact. As reported in Table R15 of our response to Reviewer 6MgT (Q6), the ViT-Small Gaussian triggers yield SSIM values of 0.94164–0.98391 and PSNR values of 35.93–39.27 dB across four datasets. Figure 2 in Appendix C (Visualization of Trigger Patterns) provides side-by-side clean/trigger-bearing image comparisons for VGG-16 and ViT-Small/16 across all four datasets, complementing these quantitative results with a qualitative assessment of both trigger types. We do not interpret either the metrics or the visualizations as proof of imperceptibility.

(2) Impact on Benign Predictions and ANN-Side Dormancy. We acknowledge that the low triggered robust accuracy (RA) of some benign ViTs indicates that Gaussian perturbations can substantially affect their predictions. However, these benign ANNs serve only as the initialization for A2S optimization and are not the models ultimately released to downstream users. The released models are the A2S-optimized ANNs, whose optimization objective explicitly encourages correct predictions on trigger-bearing inputs while enabling the attacker-specified target behavior after ANN-to-SNN conversion.

As shown in Table 1 of the original manuscript (CIFAR-10, T=4,L=6T=4,L=6), the triggered RA of ViT-Small/16 increases from 44.59% to 92.80%, while that of ViT-Base/16 increases from 18.51% to 92.78% after A2S optimization. Meanwhile, their corresponding converted SNNs achieve 100% targeted ASR. These results demonstrate that, although Gaussian perturbations can adversely affect the original benign models, A2S optimization enables the released ANNs to retain high accuracy on trigger-bearing inputs while activating the malicious target behavior only after conversion.

Overall, input-space perceptual similarity and ANN-side behavioral dormancy are distinct properties. Our additional perceptual results characterize the former, while the latter remains the central stealthiness objective of A2S-Attack.

We will revise Appendix C and the corresponding experimental descriptions before the rebuttal deadline to document Gaussian trigger generation and preprocessing, include the VGG-16 and ViT-Small/16 trigger visualizations in Figure 2 of Appendix C, incorporate the perceptual similarity analysis, and clarify the scope of our stealthiness claims.

Reviewer_6MgTOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors present a backdoor attack that is implanted in an ANN, where the backdoor remains inactive. However, when the network is converted to an SNN, the backdoor gets activated. They frame the problem as a bilevel optimization problem, with the upper level optimizing the released ANN, while the lower level simulates ANN-to-SNN conversion. The authors test on different architectures and datasets (four image datasets, four architectures, and four conversion configurations). The experimental results demonstrate very high SNN attack succes rate.

Critical strengths

Interesting idea and clear approach. I am not aware of previous work following the same idea, making the approach novel. The methodology is well discussed, and the attack framework is explained in detail.

Easy to follow. Since the paper provides a clear description of the approach and experimental validation, it is easy to assess the claims given in the paper.

Critical weaknesses

The motivation and threat model can be improved. The threat model sounds very strong and, in some elements, not realistic. This reduces the relevance of the attack.

Experimentation is not complete. While the authors provide a good set of experiments, several important aspects are not evaluated, making it more difficult to properly assess the performance.

Questions for authors

I appreciate the authors' effort on the paper. The idea is interesting, and the paper reads easily. Here are some comments that could improve the paper's quality. I think the threat model has some parts that are not realistic. First, the motivation for the setting is not clear. In which case would one convert an ANN to an SNN for use with non-neuromorphic data? What is the advantage? In which settings can this be realistic? This is not clear or motivated in the paper.

Second, the paper states that "The attacker knows the public conversion method M and its supported configuration set C". This is a strong assumption. If the defender suspects that an attack was made, could they simply try some easy things, such as using a different conversion method? A simple defense like this should be discussed.

Third, the attacker uses a surrogate calibration set, and it is not obvious how we can assume the attacker has access to one. For example, say the client wants to use an SNN for gesture detection (as in the DVS Gesture dataset). The client first gets an ANN trained on, let us say, MNIST, because there is no static version of DVS Gesture, or suppose that the dataset is private. In this case, it is hard to see how the attacker could build a realistic surrogate calibration set.

Is it possible to inject the backdoor into a model trained on CIFAR10 (static) and launch it on N-CIFAR10 (neuromorphic) after conversion? They are similar datasets, so this would be an interesting and more realistic test.

Surviving fine-tuning afterward makes a lot of sense to me. When converting to SNNs, you will want to adapt the model to your specific dataset. Since that dataset is neuromorphic, it will not be the same as, or may not even come from the same distribution as, the original static one used for training the ANN.

What if the defender converts the SNN back to an ANN and then to an SNN again? This seems straightforward, and I think the backdoor will not work in that case. The authors should evaluate it.

Trigger stealthiness evaluation is missing. How noticeable is the trigger once it is converted for the SNN? If the defender has access to a holdout test set, how different are the activations? This is discussed here: https://arxiv.org/abs/2607.28075

How many runs have been done in Table 1 and Table 2? Please report the average and standard deviation over multiple random seeds.

Ψ is used in Eq. 1, 2, and 4 but is not defined.

Post-rebuttal comment: Rebuttal addresses a number of concerns/questions I had and as such, I increased my score to 3.

13 comments

Reviewer 6MgT — Q1: Practical Motivation of ANN-to-SNN Conversion - Part 1/2

We sincerely thank the reviewer for raising this important concern. We agree that the practical motivation for ANN-to-SNN conversion on conventional data, as well as the realistic deployment scenarios underlying our threat model, deserves further clarification. We address these concerns below.

1. Practical Motivation of ANN-to-SNN Conversion

In which case would one convert an ANN to an SNN for use with non-neuromorphic data?

The primary motivation for ANN-to-SNN conversion arises from deployment requirements rather than the modality of the input data. SNNs are not restricted to processing neuromorphic data, such as event streams generated by Dynamic Vision Sensors (DVS). Conventional frame-based inputs, including RGB images, can also be processed by converted SNNs through temporal spike-based computation without requiring neuromorphic sensors.

In practice, developers may already possess high-performance pretrained ANNs for conventional vision tasks but wish to deploy these models on neuromorphic hardware to exploit its potential energy efficiency. Rather than training new SNNs from scratch, ANN-to-SNN conversion allows developers to reuse existing ANN architectures and pretrained weights while adapting their computation to spiking neurons.

This motivation is supported by existing practical deployment studies. For example, Lenz et al. [1] proposed the Quartz conversion framework and evaluated it on conventional image classification benchmarks, including MNIST, CIFAR-10, and ImageNet, with an implementation on Intel's Loihi neuromorphic processor. Chandarana et al. [2] further investigated the conversion and deployment of pretrained neural networks on neuromorphic hardware, demonstrating its practical feasibility for conventional image classification workloads.

These studies establish that ANN-to-SNN conversion provides a meaningful deployment pathway for conventional data, even when the original task does not involve neuromorphic sensors.

2. Practical Advantages of Training-Free Conversion

What is the advantage?

Training-free ANN-to-SNN conversion offers three practical advantages: (i) reduced conversion overhead, (ii) preservation of pretrained model utility, and (iii) the potential for energy-efficient inference.

First, directly training SNNs can be computationally demanding due to the non-differentiability of spike generation and the temporal dynamics involved in SNN optimization. Training-free conversion avoids additional gradient-based retraining or fine-tuning of the source ANN, allowing developers to reuse existing pretrained models with reduced training requirements. Bu et al. [3] demonstrated an efficient conversion framework with inference-scale complexity across multiple computer vision tasks, including image classification, semantic segmentation, object detection, and video classification.

Second, modern conversion methods aim to preserve the predictive performance of the source ANN while introducing spiking computation. In particular, Huang et al. [4] proposed a training-free differential coding method that achieves competitive accuracy for converted CNNs and Transformers while reducing spike activity and energy consumption.

Third, sparse, event-driven SNN computation can provide energy-efficiency benefits on suitable neuromorphic hardware. For example, Chandarana et al. [2] reported up to 27× lower power consumption and 5× lower energy consumption on Intel Loihi compared with Intel Neural Compute Stick 2 under their evaluated image classification workloads.

Importantly, these benefits depend on the conversion method, hardware platform, and deployment configuration. We do not assume that SNN deployment is universally more efficient than ANN deployment. Rather, existing evidence demonstrates that training-free ANN-to-SNN conversion provides a practically motivated deployment option under suitable hardware and application constraints.

Reviewer 6MgT — Q1: Practical Motivation of ANN-to-SNN Conversion - Part 2/2

3. Realistic Deployment Scenarios and Connection to Our Threat Model

In which settings can this be realistic?

A representative scenario involves deploying third-party pretrained ANNs on resource-constrained neuromorphic edge platforms. For example, a developer may obtain a pretrained image classification model from a third-party provider, conduct security screening on the released ANN, and subsequently apply training-free ANN-to-SNN conversion using locally selected configurations and calibration data. This workflow combines two independently motivated practices: reusing third-party pretrained models and converting existing ANNs for energy-efficient neuromorphic deployment. Prior studies have established the feasibility of ANN-to-SNN deployment on conventional image classification tasks [1–4], while model-supply-chain research has highlighted the security risks associated with third-party model distribution [5].

Crucially, this workflow creates a security boundary between the model being inspected and the model ultimately deployed. Although the released ANN may appear benign under conventional security screening, its converted SNN can exhibit different behavior due to spiking dynamics and calibration-dependent neuron parameters. This distinction introduces a security risk that cannot be ruled out through ANN-side inspection alone.

A2S-Attack specifically targets this overlooked security boundary. The attacker controls only the released ANN, while the developer independently performs conversion using private calibration samples. Our configuration-unknown setting further demonstrates that the attack can remain effective even when the exact downstream conversion configuration is not known to the attacker. To the best of our knowledge, prior SNN backdoor studies have not investigated attacks in which malicious behavior remains dormant in the released ANN and is activated specifically by ANN-to-SNN conversion. Unlike conventional backdoor transfer, our attack exploits the conversion process itself as the activation mechanism.

Therefore, our contribution is not to establish that SNN deployment is universally preferable, but to demonstrate a previously overlooked model-supply-chain vulnerability: security verification of a released ANN does not necessarily guarantee the security of its converted SNN. We will clarify this practical motivation and its security implications in the revised Introduction and Threat Model sections.

Reviewer 6MgT — Q2: Threat Model Assumptions and Simple Defense Strategies - Part 1/2

We thank the reviewer for this valuable comment. We address the attacker's knowledge assumptions and potential defenses below.

1. Threat Model Assumptions

The attacker knows the public conversion method M and its supported configuration set C. This is a strong assumption.

Setting II assumes the attacker knows the public conversion method and candidate configuration set C\mathcal C, but not the defender's eventual choice cdeploy∈Cc_{\mathrm{deploy}}\in\mathcal C. This represents an unknown deployment choice within a known candidate set, rather than arbitrary unseen configurations. We motivate this assumption as follows.

First, public ANN-to-SNN frameworks such as SpikingJelly [6, 7] document conversion procedures and configurations, making such information accessible to potential attackers without revealing private calibration samples or calibrated neuron parameters.

Second, practical configurations are constrained by accuracy–efficiency trade-offs. We evaluated T∈{2,4,6,8}T\in\{2,4,6,8\} and L∈{4,6,8,10}L\in\{4,6,8,10\}. Table R10 shows that insufficient threshold resolution degrades accuracy, while larger LL yields diminishing improvements at increased cost. At T=4T=4, increasing LL from 6 to 10 improves ACC from 91.32% to 91.87%, but raises the estimated energy proxy from 71.823 to 138.386 μJ/image. This motivates T∈{4,6}T\in\{4,6\} and L∈{6,8}L\in\{6,8\} as representative operating points.

Third, knowing the candidate set does not reveal the actual deployment configuration. In Setting II, the attacker optimizes candidate ANNs for individual configurations, evaluates their cross-configuration performance, and selects one ANN for release. Algorithm 2 provides joint refinement only when transfer is insufficient (1/16 pairs). The defender then independently selects cdeploy∈Cc_{\mathrm{deploy}}\in\mathcal C and converts the ANN using private calibration samples. The released ANN achieves high ASR across all four tested configurations and independently sampled calibration subsets, without post-release modification.

Thus, public conversion procedures, accuracy–efficiency constraints, and private deployment choices motivate our threat model. We do not assume access to private calibration data or guarantee attacks on arbitrary unseen configurations or converters.

Table R10: Accuracy-Efficiency Trade-offs under Different MTH Conversion Configurations.

Evaluated on 10,000 CIFAR-10 test images using VGG-16. Energy values are analytical estimates per image, not hardware measurements. Bold rows indicate main experimental configurations. Synaptic operation counts are estimated using SyOPs Counter (https://github.com/iCGY96/syops-counter).

TLACC (%) ↑MTH Comparisons (M) ↓Synaptic Energy (μJ) ↓Total Energy Proxy (μJ) ↓
ANN91.890.0001527.7151527.971
2481.232.27736.06538.627
2690.233.41661.06264.648
2890.424.55576.52481.136
4482.254.55536.13741.261
4691.326.83264.65071.823
4891.779.11098.171107.394
41091.8711.387127.113138.386
6482.256.83236.13743.823
6691.4810.24864.65175.411
6891.8113.66498.275112.110
61091.8517.080130.381147.290
8482.259.11036.13746.386
8691.4813.66464.65178.998
8891.7618.21998.275116.722
81091.8522.774130.384152.930
Reviewer_uBwrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies a backdoor attack on ANN-to-SNN conversion. The attacker releases an ANN that behaves normally on clean and triggered inputs, but after conversion the SNN predicts a target class for triggered inputs. A2S-Attack is a conversion-aware bilevel optimization that keeps clean ANN and SNN accuracy, suppresses the backdoor in the ANN and activates it in the SNN. Two settings are studied, one where the conversion configuration is known (Setting I) and one where it is not (Setting II). Experiments use Fashion-MNIST, CIFAR-10, CIFAR-100 and Tiny-ImageNet with VGG-16, ResNet-18, ViT-Small/16 and ViT-Base/16, and four configurations of the MTH converter. The paper also tests calibration resampling and three ANN-side detectors (Neural Cleanse, ScaleUp and BARBIE).

Critical strengths

I think the threat model is realistic. A developer can check the released ANN, even with backdoor detectors, and still deploy a converted SNN that behaves very differently. The attacker does not need to touch the conversion code or the developer's private calibration samples, which shows a clear limit of checking models only before deployment. The results are also consistent across datasets, CNNs and ViTs, and four MTH configurations (T in {4, 6} and L in {6, 8}). In Setting I the attacked ANNs have ASR of at most 0.85% while the SNNs reach about 98.93 to 100%, Setting II still gives 94.55 to 100%, and clean accuracy is mostly kept. The calibration resampling over five subsets and the detector experiments support this within the MTH pipeline.

Critical weaknesses

My main concern is novelty. Backdoors that stay dormant in a source model and become active after a downstream transformation are already known, for example after quantization (Ma et al., IEEE TDSC 2023; Egashira et al., NeurIPS 2024), after distillation with a bilevel objective (SCAR, Chen et al., NeurIPS 2025) and after fine-tuning (Li et al., AAAI 2026). SCAR is especially close in its optimization. The paper cites earlier SNN backdoor work such as Jin et al. (2026) but does not position A2S against this line of work, so the contribution seems to be showing the threat for ANN-to-SNN conversion rather than a new attack idea. A clear comparison of threat models, objectives and activation mechanisms is needed, and adapting one of these baselines would help where it makes sense. Raw ASR numbers across different transformations would not be a fair comparison on their own.

Second, I am not sure Setting II was trained as described. Algorithm 2 (App. B) starts from a Setting I ANN optimized at T=6, L=8 and then runs joint optimization over all four configurations. But in Tables 1, 2, 4 and 9, 15 of the 16 dataset and architecture pairs in Setting II have released-ANN ACC, ASR and RA identical to a Setting I checkpoint (most match T=6, L=8 and two match T=4, L=8). All ten Neural Cleanse anomaly indices for each CIFAR-10 architecture and the BARBIE statistics also match Setting I. Matching rounded metrics does not prove the weights are the same, but this much repetition needs an explanation, since configuration-robust optimization is one of the claimed contributions.

There are also some smaller issues. All experiments use the MTH converter, so it is not clear the attack works with other conversion methods, and the paper does not show what activates the backdoor (discretization, finite timestep error, threshold calibration or temporal dynamics). Detection is only tested on the released ANN, not on the converted SNN. Neural Cleanse flags some non-target classes in Setting I, and ScaleUp is shown only with confidence curves, without detection and false positive rates. The converted benign SNN with the same trigger is not reported, although some benign ANNs already change their predictions a lot when the trigger is added. The five seeds in Table 3 vary the calibration subsets, not the attack training, and App. C says there is only one attack run per model. Some key details are also missing, like how the MTH base thresholds are set from calibration samples, the surrogate gradient, input encoding and output aggregation over timesteps, whether the differential coding and threshold iteration of Huang et al. (2025) are used, and the optimization budget in each stage.

Questions for authors
  1. Beyond applying the dormant backdoor idea of Ma et al., Egashira et al., SCAR and Li et al. to a new transformation, what is the SNN-specific technical contribution of A2S?
  2. Why do 15 of the 16 Setting II released-ANN results, and the Neural Cleanse and BARBIE statistics, exactly match Setting I? How many optimizer steps are done in each stage, and can you give checkpoint identifiers or weight distances before and after Algorithm 2?
  3. Does the attack still work with a different ANN-to-SNN converter? Can you compare the attacked ANN and the SNN with a non-spiking ANN using matched activation quantization, to see if temporal dynamics are needed at all? Please also report clean accuracy, ASR and RA for the converted benign SNN with the same trigger.
  4. What happens when Neural Cleanse, ScaleUp or SNN-adapted detectors are run on the converted SNN? Could ANN/SNN disagreement on clean and triggered inputs be used to detect the attack without knowing the trigger?
11 comments

Response to Reviewer uBwr – W1 & Q1: Novelty and Comparison with Prior Work - Part 1/2

We thank the reviewer for the constructive comments regarding the novelty of A2S and its relationship to existing transformation-activated backdoor attacks.

1. Novelty and Distinction from Prior Work

"Backdoors that stay dormant in a source model and become active after a downstream transformation are already known."

We acknowledge that A2S shares similarities with existing quantization-, knowledge-distillation-, and fine-tuning-based backdoor attacks in terms of high-level attack objectives and bilevel formulations. We do not claim that the general concept of dormant backdoors or bilevel optimization is new. Rather, our contribution lies in identifying a previously unexplored security vulnerability in ANN-to-SNN conversion and addressing its distinct optimization constraints.

A New Security Scenario. To the best of our knowledge, A2S is the first systematic investigation of dormant backdoors activated specifically by ANN-to-SNN conversion. Our findings demonstrate that benign behavior in a released ANN does not guarantee the security of the SNN obtained through downstream conversion, revealing a security gap in the ANN-to-SNN deployment pipeline.

Distinct Optimization Challenges. Unlike previous transformation-based attacks, A2S must achieve reliable backdoor activation under inherited synaptic weights, calibration-dependent neuron parameters, and non-differentiable spike generation. Specifically:

  • Quantization: The source model F(⋅;ω)F(\cdot;\omega) remains benign, while the quantized model F(⋅;Q(ω))F(\cdot;\mathcal Q(\omega)) activates the backdoor by exploiting numerical changes introduced by reduced-precision representations.

  • Knowledge Distillation (e.g., SCAR): The teacher model T(⋅;ωT)T(\cdot;\omega_T) remains benign, while the student model S(⋅;ωS)S(\cdot;\omega_S) activates the backdoor through independently learned student weights ωS\omega_S.

  • Fine-tuning: The source model remains benign, while downstream parameter updates activate the dormant backdoor.

  • ANN-to-SNN: The source ANN FA(⋅;ω)F_A(\cdot;\omega) remains benign, whereas the converted SNN FS(⋅;ω,θc\*)F_S(\cdot;\omega,\theta_c^\*) inherits the same synaptic weights ω\omega, with its spiking neuron parameters θc∗\theta_c^* determined through calibration. Unlike continuous ANN activations, spike generation involves the non-differentiable Heaviside function, whose derivative is zero almost everywhere and undefined at the threshold. Consequently, A2S must jointly enforce ANN-side dormancy and SNN-side activation under shared synaptic weights, while addressing non-differentiable spike dynamics through surrogate-gradient optimization. This introduces distinct optimization constraints that do not arise in the same form in quantization-, distillation-, or fine-tuning-based attacks.

These distinctions highlight that A2S investigates a distinct conversion-dependent security problem. Its contribution is not a new general-purpose bilevel optimization paradigm, but the realization and systematic evaluation of dormant backdoors under the specific constraints of training-free ANN-to-SNN conversion.

Response to Reviewer uBwr – W1 & Q1: Novelty and Comparison with Prior Work - Part 2/2

2. Comparison with SCAR and Existing Baselines

"SCAR is especially close in its optimization."

We agree that SCAR and A2S share the high-level bilevel perspective of maintaining source-model dormancy while inducing malicious behavior after a downstream transformation. However, their lower-level processes and optimization constraints differ substantially.

Specifically, SCAR simulates knowledge distillation by optimizing an independently parameterized surrogate student model, whereas A2S simulates training-free ANN-to-SNN conversion, in which the SNN inherits the ANN's synaptic weights and its neuron parameters are determined through calibration. Furthermore, A2S must handle non-differentiable spike generation using surrogate gradients.

Regarding baseline adaptation, directly applying the original SCAR procedure would introduce downstream student training, which is absent from our threat model. Replacing SCAR's knowledge-distillation objective with ANN-to-SNN calibration would substantially alter its lower-level optimization and gradient computation, resulting in a different algorithm rather than a faithful implementation of SCAR.

To provide an empirical comparison while preserving the original SCAR attack optimization, we additionally evaluate its cross-transformation transferability. Specifically, we use SCAR's original knowledge-distillation-based attack optimization to obtain a compromised teacher ANN. Instead of performing knowledge distillation during downstream deployment, we directly convert this ANN into an SNN using the same ANN-to-SNN conversion configurations and calibration protocol as A2S.

This experiment examines whether a dormant backdoor optimized for knowledge distillation can also be activated by training-free ANN-to-SNN conversion, without modifying SCAR's original attack optimization.

Table R1: Cross-Transformation Transferability of SCAR from Knowledge Distillation to ANN-to-SNN Conversion.

Results for VGG-16 and ResNet-18. Teacher ANN and Student ANN refer to SCAR's original source and distilled models, respectively. Teacher → SNN denotes direct conversion of SCAR's compromised teacher ANN using the same MTH calibration protocol as A2S. ACC and ASR are reported in percentages (%); T and L denote the SNN conversion configuration.

Teacher ArchitectureEvaluated ModelT/LACC (%)ASR (%)
VGG-16Teacher ANN91.280.31
Student ANN90.2595.43
Teacher → SNN4/690.411.24
Teacher → SNN4/890.241.31
Teacher → SNN6/690.411.26
Teacher → SNN6/890.231.31
ResNet-18Teacher ANN91.150.57
Student ANN87.8096.06
Teacher → SNN4/690.370.31
Teacher → SNN4/890.240.53
Teacher → SNN6/690.380.36
Teacher → SNN6/890.240.54

As shown in Table R1, SCAR successfully activates its backdoor through knowledge distillation, achieving student ANN ASRs of 95.43% and 96.06% for VGG-16 and ResNet-18, respectively. However, directly converting the compromised teacher ANNs into SNNs yields ASRs of only 0.31%–1.31% across all evaluated configurations, while maintaining approximately 90% clean accuracy.

These findings demonstrate limited cross-transformation transferability under the evaluated settings: a backdoor optimized for knowledge distillation does not necessarily activate after training-free ANN-to-SNN conversion. This supports the motivation for A2S's conversion-specific optimization, rather than relying on the direct transfer of an existing transformation-activated backdoor.

Response to Reviewer uBwr – W2 & Q2: Clarification of Setting II Training - Part 1/2

We thank the reviewer for this careful observation. We acknowledge that the original manuscript did not sufficiently distinguish between checkpoint reuse and additional joint optimization in Setting II. We clarify the checkpoint selection strategy, the role of Algorithm 2, and the supporting evidence below.

1. Identical ANN Metrics Between Settings I and II

"But in Tables 1, 2, 4 and 9, 15 of the 16 dataset and architecture pairs in Setting II have released-ANN ACC, ASR and RA identical to a Setting I checkpoint (most match T=6, L=8 and two match T=4, L=8). All ten Neural Cleanse anomaly indices for each CIFAR-10 architecture and the BARBIE statistics also match Setting I."

During our experiments, we observed that ANNs optimized for a single conversion configuration in Setting I often exhibited strong cross-configuration transferability. For each dataset–architecture pair, we evaluated the available Setting I checkpoints across all four configurations and selected the checkpoint with the strongest cross-configuration attack performance as the initialization for Setting II.

We adopted a transfer-first, refinement-if-needed strategy: checkpoints achieving satisfactory attack effectiveness across configurations were directly reused; otherwise, Algorithm 2 was applied for joint refinement. Consequently, 15 of 16 pairs required no additional updates, while TinyImageNet–ViT-Base/16 underwent 100-step joint refinement.

This explains the identical released-ANN ACC, ASR, RA, Neural Cleanse anomaly indices, and BARBIE statistics: they were obtained from the same reused checkpoints rather than independently optimized models.

We will revise the manuscript before the rebuttal deadline to document checkpoint selection, provenance, and the distinction between direct reuse and joint refinement.

2. Necessity and Effectiveness of Algorithm 2

"Matching rounded metrics does not prove the weights are the same, but this much repetition needs an explanation, since configuration-robust optimization is one of the claimed contributions."

Importantly, cross-configuration transfer is not guaranteed. For TinyImageNet–ViT-Base/16, the same initial Setting I checkpoint exhibited near-zero ASR under two configurations despite high ASR under the other two, demonstrating the limitations of single-configuration optimization.

We therefore applied 100 additional optimizer updates using Algorithm 2, jointly optimizing across all four configurations.

Table R2: Comparison of SNN ACC and ASR (%) for TinyImageNet–ViT-Base/16 before and after joint optimization.

ConfigurationBefore ACCBefore ASRAfter ACCAfter ASR
T=4, L=678.490.0588.49100.00
T=4, L=888.9399.1289.0199.93
T=6, L=678.960.0388.96100.00
T=6, L=888.9199.4688.8299.05

As shown in Table R2, joint refinement increased ASR from 0.05% and 0.03% to 100.00% under the two failing configurations, while improving clean accuracy by approximately 10 percentage points in both cases. ASR remained above 99% under the other configurations, yielding a minimum ASR of 99.05% across all four configurations.

These results demonstrate that Algorithm 2 can recover attack effectiveness when direct checkpoint transfer fails, while preserving high performance under configurations where transfer already succeeds.

Therefore, the principal empirical finding in Setting II is cross-configuration transferability of selected Setting I checkpoints (15/16 pairs). Algorithm 2 provides a conditional refinement mechanism, whose benefit is demonstrated on TinyImageNet–ViT-Base/16. We do not claim that all 16 models underwent joint configuration-robust optimization.

Comments on the paper

1 comment

Summary of Our Rebuttal and Additional Evidence

Dear Area Chair,

Thank you very much for your time and effort in handling our submission.

As the discussion period approaches its end, we would like to summarize our rebuttal for your final assessment. We have addressed the reviewers' concerns through extensive additional experiments (Tables R1–R20). The key findings are highlighted below.

1. Novelty and Distinction from Prior Work (uBwr: W1/Q1; mtCo: W2).

A2S-Attack identifies a previously unexplored security vulnerability in ANN-to-SNN conversion, rather than proposing a general-purpose bilevel optimization paradigm. Unlike existing transformation-activated backdoors, A2S must maintain ANN-side dormancy and SNN-side activation under shared weights, calibration-dependent neuron parameters, and non-differentiable spiking dynamics. Our comparison with SCAR (Table R1) shows that its distillation-activated backdoor achieves only 0.31%–1.31% ASR after ANN-to-SNN conversion, supporting conversion-specific optimization.

2. Generalizability and Practical Relevance (uBwr: Q3; 6MgT: Q1/Q3/Q4; mtCo: W1).

We extended our evaluation to an alternative training-free IF-based converter, achieving 98.92%–100% ASR (Table R4). On neuromorphic CIFAR10-DVS data, ASR increases from 0.05% in the released ANN to 99.2% after conversion (Table R13). Cross-distribution calibration experiments also demonstrate consistently high ASR when CIFAR-100 samples calibrate CIFAR-10 models (Table R12). These results strengthen the practical relevance of A2S within our threat model.

3. Attack Mechanism and Optimization Analysis (uBwr: Q3; mtCo: W4).

Comparisons among attacked ANNs, activation-quantized ANNs, and converted SNNs (Table R5), together with timestep analysis (Table R6), show that static quantization alone cannot reproduce the near-100% SNN ASR, highlighting conversion-specific computation. Loss ablations (Table R20) further validate the importance of ANN dormancy and SNN activation objectives.

4. Configuration Robustness and Reproducibility (uBwr: W2/Q2; 6MgT: Q2/Q7; mtCo: W1/W5).

We clarified that 15 of 16 dataset–architecture pairs achieve cross-configuration transfer through checkpoint selection, while the remaining pair requires joint refinement. For TinyImageNet–ViT-Base/16, Algorithm 2 improves ASR from nearly 0% to 100% under two configurations (Tables R2–R3). Three independent optimization runs across CNNs and Transformers further demonstrate high SNN ASR with low variance (Tables R17–R18).

Regarding the candidate configuration set, public frameworks such as SpikingJelly document conversion procedures and configurations, while our accuracy–efficiency analysis (Table R10) motivates practical configuration choices. Thus, our configuration-unknown setting considers an unknown deployment choice within a motivated candidate set, not arbitrary unseen configurations. Our primary contribution is uncovering the ANN-to-SNN conversion vulnerability; robustness to arbitrary unknown configurations remains future work.

5. Trigger Stealthiness, Security Evaluation, and Defense Analysis (uBwr: Q4; 6MgT: Q2/Q6; mtCo: W3/W5).

We quantified trigger perceptibility across four datasets, obtaining SSIM of 0.9371–0.9839 and PSNR of 35.81–39.27 dB (Table R15). Visualizations of CNN patch and ViT Gaussian triggers (Figure 2, Appendix C) complement these results. Activation-space analysis and Spectral Signatures detection (Table R16) reveal distinct behaviors between released ANNs and converted SNNs. We also evaluated Neural Cleanse and ScaleUp on SNNs (Table R8), ANN/SNN prediction disagreement (Table R9), and converter-switching defenses (Table R11), clarifying detectability and defense limitations.

Together, these results substantially strengthen our paper and address major concerns regarding novelty, generalizability, mechanisms, reproducibility, and security implications.

We recognize that no single study can exhaustively address every possible extension, alternative setting, or defense strategy. Nevertheless, we have addressed the reviewers' concrete concerns through targeted experiments and analyses. We respectfully hope the paper will be evaluated based on its central contribution, supporting evidence, and the extent to which the original concerns have been addressed, while distinguishing these from future research directions.

We sincerely appreciate Reviewer 6MgT's follow-up, recognition of our additional experiments, and increased score. However, we have not received follow-up feedback from the other two reviewers, limiting our opportunity to determine whether their concerns have been resolved. We would therefore particularly appreciate the Area Chair's careful consideration of our rebuttal and additional findings when assessing whether the original concerns remain valid in light of the new evidence.

Thank you again for your time and consideration.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.