acceptodds
Under review as a conference paper at ICLR 2027

Stealth Is a Relation, Not a Property: How Event Representations Create Blind Spots for Timing Attacks in Event-Based Perception

Abstract

An event camera produces an asynchronous stream, but what is visible in that stream depends on how a downstream consumer, such as a model or detector, processes time. The same timestamp change may leave a coarse temporal representation unchanged while changing the response of a model that preserves finer timing. This creates blind spots that timing attacks can exploit. We characterize this dependence as observer-relative stealth. For recorded event streams, retiming an event within its protected accumulation window leaves the accumulated integer tensor exactly unchanged. Any consumer that receives only this tensor, therefore, sees the same input before and after the attack. We use this exact blind space to construct Null, a gradient-guided timestamp-retiming attack, and define to measure attack success while bounding the change visible to observer . Across five event-vision datasets, the exact blind space contains retimings that remain exactly hidden from the protected observer while causing victim failure. On DVS Gesture at a 10% event budget, Null reaches ASR on ConvSNN and on a GRU while preserving the protected tensor exactly. On DailyDVS-200, a protocol-scale Multi-View Fusion Network variant reaches exact-null ASR, compared with for its matched control. In a five-attack comparison, Null is the only method with nonzero attack success at exact observer equality, reaching on DVS Gesture and on DailyDVS-200. We also search the same exact blind space with an independently implemented constrained projected-gradient optimizer, C-PGD. At matched victim-gradient evaluations, C-PGD reaches ASR on DVS Gesture and on DailyDVS-200, again with exact protected equality. Both optimization procedures find retimings with high attack success within the exact same blind space. We then change the temporal observer and test the same attacks. Perturbations that are exactly hidden from the protected observer become visible under shifted, finer, overlapping, and randomized temporal views. Adding observer constraints reduces the real-valued blind-space fraction from to to , while DVS ConvSNN ASR falls from to to . These results show that stealth is not a property of the perturbation alone. It depends on the temporal information available to the observer.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
event-based vision, event cameras, adversarial robustness, temporal attacks, spiking neural networks, neuromorphic vision, temporal representations, adversarial machine learning

Reviews

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

This paper studies timestamp attacks on event streams that preserve a coarse temporal representation exactly while causing errors in models using finer timing. It introduces Null, validates the vulnerability with an independent constrained optimizer, and shows that changing temporal observers can expose previously invisible perturbations.

Critical strengths
  1. Clearly distinguishes attack success, representation-level invisibility, and detector evasion.

  2. Evaluates multiple datasets and architectures that supports their performance.

Critical weaknesses
  1. The core invariance follows directly from within-window count preservation, and related work already studies count-preserving timestamp manipulation. The main contribution is limited.

  2. Incomplete baseline comparison: Existing attacks are evaluated against an exact-invariance constraint they were not designed to satisfy. Adapting them to the same feasible set would better establish the incremental contribution.

  3. Some models have low clean accuracy or use custom implementations, resulting in weak demonstration.

  4. Some related references [1, 2] are missing.

  5. The central idea is clear, but the presentation is repetitive, numerically dense, and places important qualifications in the appendix.

[1] Wang R, Guo Q, Li H, et al. Event trojan: Asynchronous event-based backdoor attacks[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 315-332.

[2] Gao S, He W, Chang C H, et al. Spatiotemporal Trigger on Event Data for Backdoor Attack on Spiking Neural Networks[C]//2025 IEEE 7th International Conference on Artificial Intelligence Circuits and Systems (AICAS). IEEE, 2025: 1-5.

Questions for authors

No further questions.

6 comments
AuthorsOpenReview ↗

Response to Reviewer mrHv [1/4]

We thank the reviewer for the careful review and for recognizing the distinction between attack success, representation invisibility, and detector evasion. The comments motivated additional experiments and a clearer presentation of our contributions.

W1. Contribution beyond count-preserving retiming. We agree that count-preserving retiming has precedents, including Yu et al., Time Is All It Takes: Spike-Retiming Attacks on Event-Driven Spiking Neural Networks (ICLR 2026), and Riaño et al., Temporal Poisoning: Clean-Label Backdoors via Event Redistribution in SNNs (arXiv 2026), both discussed in our submission. Our contribution is to formulate stealth relative to a specified observer and establish the security consequences of this relationship. For a clean event stream XX, attacked stream X′X', and observer RR, exact stealth holds when R(X′)=R(X)R(X') = R(X). We characterize the observer's temporal blind space, demonstrate that it contains adversarially effective perturbations, and examine how changing the observer changes their visibility.

First, we demonstrate observer-dependent stealth using fixed attacks. On DVS Gesture, we generated phase-0-preserving attacks and evaluated the same 709 attacked seed-clip streams under three observers. Each coarse window accumulates eight of 160 fine bins. Phases 2 and 4 shift the accumulation boundaries by two and four bins.

Same 709 attacked streamsPhase 0 (protected)Phase 2Phase 4
Representation visibilityInvisible in 709/709Visible in 709/709Visible in 709/709
Normalized count difference (DAD_{A}, % of clean count)0%11.6%13.7%

The attacked streams and victim predictions remain fixed. Changing only the observer changes whether the same perturbations are visible. This directly demonstrates why stealth is a relation rather than a property.

Second, we connect the observer's blind space to attack feasibility and effectiveness. We independently optimized attacks under increasingly restrictive observer configurations.

Protected phases00, 20, 2, 4
Blind-space fraction87.5%75.0%62.5%
Legal fine-bin moves1,120640320
ConvSNN SC-ASR(0)81.70%65.18%43.05%

SC-ASR(0) measures attack success under exact preservation of every protected observer. Values are three-seed means.

Even when all three observers are preserved exactly, attacks successfully change 43.05% of initially correct predictions. Together, these experiments establish observer-dependent visibility and effective attacks within intersecting blind spaces. We have clarified these contributions in Sections 1 and 3 and added the observer analyses to Section 5.

Continued in Part 2/4 below.

AuthorsOpenReview ↗

Response to Reviewer mrHv [2/4]

W2. Constraint-matched comparison with existing attacks. Our original comparison examined whether attacks optimized for their native objectives also remained invisible to a common temporal observer. This characterizes their observer-relative behavior. Following the reviewer's suggestion, we extended the evaluation to compare attacks explicitly optimized under the same preservation requirement.

We adapted the published Yu PIL-L0L_{0} retiming attack of Yu et al., Time Is All It Takes (ICLR 2026), to enforce exact coarse-observer equality during attack generation. We compared the adapted Yu attack with Null and C-PGD using three DVS Gesture ConvSNN seeds, 709 initially correct seed-clip evaluations, and matched victim-gradient counts of 3, 6, and 12. All final attacks satisfy the same declared observer equality and original-event modification budget, B=round⁡(0.1N)B=\operatorname{round}(0.1N).

MethodG = 3G = 6G = 12
Null81.41 ± 4.99%89.01 ± 2.45%91.55 ± 2.15%
C-PGD84.50 ± 2.52%88.73 ± 2.70%90.42 ± 1.83%
Adapted Yu PIL-L0L_{0}6.92 ± 1.54%46.70 ± 3.62%87.61 ± 3.07%

Three-seed mean ± SD of SC-ASR(0). Every final attack satisfies exact observer equality and the declared event-edit budget.

Paired advantage over adapted YuG = 3G = 6G = 12
Null+74.47 pp+42.31 pp+3.95 pp
C-PGD+77.57 pp+42.03 pp+2.82 pp

Clip-clustered 95% confidence intervals at G3 are [69.83, 78.74] for Null and [73.15, 81.79] for C-PGD. At G12, they are [2.22, 5.96] and [1.42, 4.37].

The constraint-matched comparison shows that the observer's blind space is exploitable by an adapted published method, which achieves 87.61% SC-ASR(0) at G12. Our constrained search achieves high success with fewer gradient evaluations, with Null reaching 81.41% versus 6.92% for adapted Yu at G3. Null and C-PGD achieve higher success at all three evaluated budgets.

Yu is the closest published timestamp-retiming baseline. The other evaluated baselines use different event-edit mechanisms, so adapting them would require changing their original attack operations. We have added the adapted-Yu comparison alongside the original observer-relative evaluation in Section 5.

Continued in Part 3/4 below.

AuthorsOpenReview ↗

Response to Reviewer mrHv [3/4]

W3. Clean accuracy and official checkpoint validation. We agree that accurate models and independently released checkpoints strengthen the security evidence. Our submitted DVS Gesture evaluation already includes high-accuracy models.

ModelClean accuracyExact-observer result
Event Transformer93.69 ± 0.95%51.0% SC-ASR(0) vs. 1.1% random control
CoarseFrameFormer, protected consumer93.81 ± 0.22%Bit-identical protected logits

The Event Transformer comparison uses a displacement-matched random control under the 10% original-event modification budget. Lower-accuracy variants were identified as supporting evaluations.

We trained the original architecture variants using shared dataset-specific preprocessing to enable controlled comparisons across models and temporal observers. Following the reviewer's suggestion, we additionally evaluated the officially released Event-SSM small checkpoint, retaining its pretrained weights and native asynchronous-event processing without retraining. We obtained 98.11% clean accuracy (259/264 recordings).

Event-SSM clean accuracyNull G3 SC-ASR(0)Null G12 SC-ASR(0)Protected observer exact
98.11% (259/264)19.31% (50/259)30.50% (79/259)100%

At G3, four-query black-box search achieves 1.16% (3/259), giving Null a paired advantage of +18.15 percentage points (95% CI [13.51, 22.78]). These results establish successful exactly observer-preserving attacks on an independently released, 98%-accurate model processing native asynchronous events.

We also realized the submitted ConvSNN attacks as timestamp modifications on recorded raw DVS Gesture streams, covering all 709 initially correct seed-clip evaluations.

Raw-to-tensor parityCoarse representation exactProtected logits identicalFine-time misclassificationOutcome agreement
709/709709/709709/709578/709 (81.5%)709/709

(The replay reproduces all submitted attack outcomes, including 578 flips and 131 nonflips. Three-seed mean 81.56 ± 5.81%, pooled rate 81.5%.)

The official checkpoint extends the validation to a high-accuracy released model, while the raw-event replay confirms the submitted attack outcomes on recorded event streams. The submitted Null result (81.56%), matched-gradient comparison (81.41%), and observer-intersection experiment (81.70%) are separate runs and are reported accordingly. We have included the additional validation in Section 5.

W4. Missing related work. We thank the reviewer for identifying Wang et al., Event Trojan (ECCV 2024), and Gao et al., Spatiotemporal Trigger on Event Data for Backdoor Attack on Spiking Neural Networks (AICAS 2025). We have expanded the Related Work discussion in Section 2 to include both studies and clarify their relationship to our work, and we added them to the event-security comparison table in Appendix B.1.

Our work addresses a distinct security problem. Event Trojan and Gao et al. study backdoor attacks, which require influence over training to implant a trigger that a compromised model later responds to. Our attack instead targets fixed, unmodified victim models at test time and only retimes existing events, so event identities, locations, polarities and counts are unchanged. Stealth in backdoor work means the trigger is inconspicuous to a human or a defense. Here it is an exact, measurable relation, R(X′)=R(X)R(X')=R(X) for a named observer RR, and we show that the same perturbation can be invisible to one observer and visible to another.

Continued in Part 4/4 below.

Reviewer_1UXwOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies stealth relative to a specified temporal observer. Timestamp changes within a protected accumulation window preserve its event-count tensor exactly, while potentially changing predictions of models that retain finer timing. The authors introduce Null, a gradient-guided retiming attack, and independently search the same feasible set using C-PGD. They distinguish raw attack success, observer-constrained success, and detector evasion. Experiments across event-vision datasets and an auditory benchmark demonstrate successful exact-null attacks; changing or intersecting temporal observers exposes some perturbations and reduces attack success for some architectures.

Critical strengths
  • The distinction between observer-level invisibility and detector evasion is valuable. Exact integer equality, protected-consumer checks, displacement-matched controls, and the Null–C-PGD comparison provide complementary evidence that successful attacks exist inside the declared blind space.
Critical weaknesses
  • Some victims have low clean accuracy, and the protocol-scale architecture variants are trained under the authors’ preprocessing rather than evaluated as official benchmark checkpoints. The paper discloses these limitations, but stronger-checkpoint validation and a concrete deployment example would strengthen the security significance of preserving the selected observer exactly
Questions for authors
  • What concrete deployment motivates protecting the specified coarse observer while attacking a separate fine-time consumer?
  • Where would the attacker obtain timestamp-modification access?
6 comments
AuthorsOpenReview ↗

Response to Reviewer 1UXw [1/2]

We thank the reviewer for the careful review and for recognizing the importance of exact observer preservation. We conducted additional experiments to strengthen the checkpoint validation and deployment evidence.

W1. Clean accuracy and official checkpoint validation. We agree that observer-preserving attacks should be demonstrated on accurate models. Our submitted DVS Gesture evaluation includes:

ModelClean accuracyAttack evidence
ConvSNN89.5% (709/792)Main victim in submitted experiments
Event Transformer93.69 ± 0.95%51.0% SC-ASR(0) vs. 1.1% displacement-matched random control
CoarseFrameFormer93.81 ± 0.22%Bit-identical protected logits

SC-ASR(0) measures attack success under exact observer equality. The Event Transformer comparison uses the declared 10% original-event edit budget. Lower-accuracy variants were identified as supporting evaluations in the submitted paper.

The original architecture variants used shared dataset-specific preprocessing for controlled comparisons across victim architectures and temporal observers. As per the reviewer’s suggestion, we additionally evaluated the officially released Event-SSM small checkpoint (official implementation: https://github.com/Efficient-Scalable-Machine-Learning/event-ssm) using its pretrained weights and native asynchronous-event pipeline, without retraining.

Clean accuracyNull G3 SC-ASR(0)Null G12 SC-ASR(0)Black-box Q4 SC-ASR(0)Null G3 advantage over Q4Exact observer preservation
98.11% (259/264)19.31% (50/259)30.50% (79/259)1.16% (3/259)+18.15 pp [13.51, 22.78]100%

Official Event-SSM results. Event-edit budget: B=round⁡(0.1N)B=\operatorname{round}(0.1N). G is gradient evaluations, Q is forward queries, and brackets show the paired 95% confidence interval. Null G3 and black-box Q4 each use four forward evaluations, while Null additionally uses three backward passes.

These results establish successful exactly observer-preserving attacks on an independently released, 98%-accurate model processing native asynchronous events.

Q1. Concrete two-consumer deployment. We consider a shared post-sensor event-processing pipeline where two software consumers receive the same timestamped event stream. A coarse-frame consumer accumulates events for monitoring or logging, while a fine-time consumer uses precise event timing for recognition. Both representation types are established in event-based perception. Wang et al., DailyDVS-200 (ECCV 2024, Table 2), evaluate five frame-based methods using 500 ms event accumulation alongside token- and spike-based methods.

We validated this scenario through raw-event replay on DVS Gesture. We reconstructed the submitted Null attacks as timestamp-only modifications of recorded event streams and replayed them through the frozen CoarseFrameFormer and ConvSNN consumers. The evaluation covers 709 initially correct seed-clip cases across three seeds, representing 244 distinct recordings. The attacks modify 9.52% to 9.94% of event timestamps while preserving event identities, coordinates, polarities, and counts.

Raw-to-tensor parityExact coarse representationIdentical protected logitsConvSNN misclassificationAgreement with submitted outcomes
709/709709/709709/709578/709 (81.5%)709/709

All 709 submitted outcomes are reproduced, including 578 successful and 131 unsuccessful cases. Per-seed rates match the submission (three-seed mean 81.56 ± 5.81%; pooled 81.5%).

The replay demonstrates the two-consumer security scenario directly. The coarse consumer receives an identical representation and produces identical logits in every case, while the fine-time consumer is misled in 578 cases.

Continued in Part 2/2 below.

AuthorsOpenReview ↗

Response to Reviewer 1UXw [2/2]

Q2. Post-sensor timestamp-modification access. Our threat model assumes that the attacker can modify event timestamps after sensor acquisition and before downstream processing, for example through a compromised event driver or intermediate event buffer. The attacker knows the protected observer's accumulation schedule and, in the evaluated white-box setting, has gradient access to the fine-time victim.

The attack changes only timestamps while preserving event identities, coordinates, polarities, counts, and model parameters. The raw-event replay demonstrates that the coarse consumer's representation and logits remain identical while the fine-time consumer's predictions change.

We hope the official-checkpoint evaluation and raw-event validation address the reviewer's questions about checkpoint quality, deployment relevance, and attacker access. We thank the reviewer again for the insightful review. Please let us know if you have any further comments or questions. We look forward to a fruitful discussion.

AuthorsOpenReview ↗

Feedback Request to Reviewer 1UXw

We thank the reviewer again. Please let us know if you have any further feedback or queries. We look forward to continuing the discussion.

Reviewer_i3MbOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper argues that the asynchronous nature of event camera streams and the varying way in which time is processed by downstream models create vulnerable blind spots. This dependence is characterised and leveraged as observer-relative stealth, and a gradient-guided timestamp-retiming Null attack, is proposed to retime events within their protected accumulation window.

Critical strengths

The paper provides empirical results showing that stealth depends not only on perturbations but also on the temporal information available to the observer. It also empirically demonstrates that effective Null attacks can be constructed.

Critical weaknesses
  1. The title emphasises that "stealth is a relation, not a property", which sounds attractive but is not fully explained or supported. It explains that "stealth is not a property of the perturbation alone", which is weaker than just "not a property" in the title. Besides, it does not explain what the word "relation" means. The later description says "(stealth) depends on the temporal information available to the observer", but this does not establish a connection between "relation" and "availability of temporal information". The paper should explain why it highlights stealth as a relation; at least explain whom it is a relation between. Although the current title sounds groundbreaking, it is beyond the evidence that the paper actually established.

  2. Section 3 does not provide much new information. Proposition 1 is basically saying that if R(X′)=R(X)R(X') = R(X), then R(X′)R(X') and R(X)R(X) are the same. Similarly, Eq. 3 is an equation derived from the definition, and Proposition 3 is saying that if we define the blind space as the null space, then the intersection of blind spaces is the intersection of null spaces. Frankly, these claims are technically always true, but there is little new information we can learn.

  3. The abstract and introduction focus too much on listing numerical results and hide the really important argument within them. The results in Table 2 are the most direct evidence supporting the paper's main claim, but they are placed at the end of the abstract and introduction, behind many other supplementary results. The paper should make its main claim and support clear.

Questions for authors

Please address the concerns mentioned in the critical weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer i3Mb [1/2]

We thank the reviewer for the careful review and for recognizing that stealth depends on the temporal information available to the observer. We address each point below with additional experiments and manuscript clarifications.

W1. Why is stealth a relation rather than a property, and between what? We agree that this distinction should be explicit. Stealth is a relation among three objects: the clean event stream XX, the perturbed stream X′X', and the observer RR. Definition 1 (Section 3) states that exact stealth holds when R(X′)=R(X)R(X') = R(X). Thus, the same perturbation can be invisible to one observer and visible to another, depending on the temporal information each observer preserves. Coarse accumulation retains window counts but discards within-window timing. Retimings within a window are therefore invisible to that observer, while shifted accumulation boundaries can reveal them.

We tested this directly using fixed attacks on DVS Gesture. We generated phase-0-preserving attacks and evaluated the same 709 attacked seed-clip streams under three observers. Each coarse window accumulates eight of 160 fine temporal bins. Phases 2 and 4 shift the boundaries by two and four bins, respectively.

Same 709 attacked streamsPhase 0 (protected)Phase 2Phase 4
Representation visibilityInvisible in 709/709Visible in 709/709Visible in 709/709
Normalized count difference (DAD_{A}, % of clean count)0%11.6%13.7%

Visibility means DA=∥Aϕ(X′)−Aϕ(X)∥1∥Aϕ(X)∥1>0D_A=\frac{\lVert A_\phi(X')-A_\phi(X)\rVert_1}{\lVert A_\phi(X)\rVert_1}>0, where AϕA_\phi accumulates eight consecutive fine bins per window using offset ϕ\phi.

The attacked streams and victim predictions remain fixed. Changing only the observer changes whether the perturbations are visible. This directly establishes our central claim that stealth is a relation rather than a property. We have clarified this interpretation in the Introduction (Section 1) and Definition 1 (Section 3).

W2. Section 3 provides limited new information because its mathematical statements are elementary. We agree that Proposition 1, Equation 3, and Proposition 3 follow from representation equality and standard null-space properties. Their role is to construct and verify exactly observer-preserving attacks, rather than serve as independent theoretical contributions.

The preservation condition enables timestamp edits satisfying exact representation equality (DA=0D_{A} = 0). The intersection characterization identifies retimings that preserve multiple observers simultaneously and determines the blind-space fraction and legal moves before attack optimization.

Our contribution is to establish whether these mathematically admissible retimings are adversarially effective. We tested this by independently optimizing ConvSNN attacks under three observer configurations on DVS Gesture.

Protected observersBlind-space fractionLegal fine-bin movesMean timestamp shiftConvSNN SC-ASR(0)
Phase 087.5%1,120181 ms81.70%
Phases 0, 275.0%640128 ms65.18%
Phases 0, 2, 462.5%32075 ms43.05%

Three-seed means over 709 initially correct seed-clip evaluations per configuration. SC-ASR(0) measures successful attacks while every protected representation remains exactly unchanged.

The results establish two findings. First, adding observers reduces the common blind space and available timestamp retimings. Second, 43.05% attack success remains achievable even when all three observers are preserved exactly.

The mathematical identities characterize which perturbations are admissible, while our experiments establish that these blind spaces contain effective adversarial perturbations. We have clarified this distinction in Section 3 and added the observer-intersection results to Section 5.

Continued in Part 2/2 below.

AuthorsOpenReview ↗

Response to Reviewer i3Mb [2/2]

W3. The abstract and introduction obscure the central argument behind numerical results. We agree that the central claim and its direct evidence should appear before the broader experimental results. We thank the reviewer for identifying Table 2 of the original submission as the most direct evidence supporting our argument.

We have revised the Abstract and Introduction (Section 1) to emphasize the observer-relative formulation, explain how coarse temporal accumulation creates blind spaces, and introduce the fixed-attack evidence demonstrating observer-dependent visibility.

We have also moved the original cross-observer comparison to Table 1 at the beginning of Section 5, extending it with the fixed-stream exactness results reported in W1. The subsequent evaluation presents the constrained-baseline comparison, official-checkpoint validation, and raw-event replay as complementary evidence of attack effectiveness and security significance.

We thank the reviewer again for the insightful review. Please let us know if you have any further comments or questions. We look forward to a fruitful discussion.

AuthorsOpenReview ↗

Feedback Request to Reviewer i3Mb

We thank the reviewer again. Please let us know if you have any further feedback or queries. We look forward to continuing the discussion.

Dear Reviewer i3Mb,

Thanks again for your constructive feedback. We have tried to address your concerns and updated the paper as per your suggestion.

With just few hour left within the discussion window, please let us know if you have any further thoughts.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Request for feedback

Dear Reviewers,

Thank you for your time and thoughtful evaluations. As the discussion period approaches its end, we would sincerely appreciate it if you could kindly review our responses and share any further feedback. We hope our responses have addressed your questions and concerns, and we would be grateful if you could take them into consideration when reassessing your evaluations.

Thank you once again for your valuable 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.