acceptodds
Under review as a conference paper at ICLR 2027

EventCert: Certified Robustness of Event-Based Classifiers to Structured Deletion

Abstract

Event cameras record the changes in a scene as a stream of events, which is turned into a tensor so that standard models can process it efficiently. In practice, packet loss, sensor faults, or occlusion can remove whole blocks of this tensor at once. In a traditional camera, such a loss does limited damage because neighbouring pixels carry similar information. Such redundancy is almost absent in event streams. As a result, a lost block is lost for good and can change a classifier’s prediction, posing serious risks in safety-critical applications such as robotics and driving. It is therefore essential to prove how many blocks can be removed while the prediction is guaranteed to stay the same, a question existing certification methods have largely left unexplored. We close this gap with EventCert, a method that proves, with high probability, how many blocks each prediction can lose without changing. It mirrors how events are deleted and certifies how many blocks an adversary may remove. Because only blocks that contain events can change the prediction, it turns this worst-case certificate into the first exact guarantees for random and bursty channel loss, and for an adversary and a lossy channel acting together, all from a single certification run. On five event-camera and neuromorphic-audio benchmarks, we compare EventCert with three existing certification methods: Gaussian smoothing, randomized ablation, and random deletion of single events. EventCert certifies 3.8–5.6× more deleted blocks than the best of them, while the other two certify almost none. On N-MNIST, 89% of test inputs stay certified even when an adversary removes any 8 blocks. While standard metrics only tell us how often a model is right, EventCert tells us, for each prediction, whether it will stay the same when part of the data is lost. By telling when an event-camera decision can still be trusted and when it cannot, EventCert can make self-driving cars safer, drones more reliable, and autonomous robots more effective.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
certified robustness, randomized smoothing, event-based vision, neuromorphic computing, event streams, structured deletion, spatio-temporal robustness, deletion robustness, sparse sensing, physical sensor loss, active-group certification, robust machine learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_d7o2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a method for certifying the robustness of classifiers for structured event-based data, such as those obtained from event cameras, against group deletion. Specifically, the authors assume that the input tensor is divided into groups and use a Randomized Smoothing framework extended to group deletion to derive a worst-case certified radius against group deletions. Also, by using the fact that deleting inactive groups containing no events does not change the input, the authors extend the worst-case certificate to probabilistic guarantees against random loss, bursty loss, and hybrid threats combining adversarial deletion and random loss. The authors also show that the proposed method gives larger certified radii compared with existing methods on five benchmarks.

Critical strengths
  1. The paper proposes an RS-based certification framework that accounts for group-level deletion in structured data, such as data obtained from event cameras. The framework also provides guarantees against random and bursty losses that may occur in practical systems.
Critical weaknesses
  1. Several extensions of RS that account for data structures, such as graphs, point clouds, image patches, and sequences, have already been proposed. However, the novelty of the proposed method relative to these existing approaches is not sufficiently clear. Experimental comparisons with related structured RS methods are also absent.
  2. The accuracy–robustness trade-off controlled by p_keep is not sufficiently investigated. Although Figure 1 shows how certified accuracy varies with p_keep, the relationship between smoothed classification accuracy and certified radius is not directly evaluated.
Questions for authors
  1. What is the specific theoretical novelty of the proposed method compared with existing structured RS frameworks, such as Hierarchical Randomized Smoothing (Scholten et al., 2023)? In particular, could the worst-case and physical-loss certificates presented in this paper be derived by applying existing frameworks to SST groups? If not, please clarify what prevents such a derivation. Would it also be possible to include theoretical or experimental comparisons with these closely related methods?
  2. Could the authors directly examine the relationship between smoothed classification accuracy and average certified radius (ACR) as p_keep varies across the datasets? In particular, it would be helpful to clarify where the current choice of p_keep, selected by maximizing ACR, lies in the accuracy–robustness trade-off.
5 comments
AuthorsOpenReview ↗

Novelty vs. structured RS, and the full accuracy–ACR trade-off

Thank you for the clear summary. Both questions are answered below, including the full trade-off data you asked for.

Weakness 1 and Question 1: How is EventCert new compared with structured RS such as HRS? Can existing frameworks derive its certificates?

The worst-case certificate, yes. The paper does not claim it as new (l. 59–68). For Bernoulli group deletion, the likelihood ratio between the smoothed DS(X)\mathcal{D}_S(X) and XX lies in 0,qk−1\\{0,q_k^{-1}\\} with qk=(1−pkeep)kq_k=(1-p_{\mathrm{keep}})^k. The Neyman–Pearson lemma then gives exactly Eq. 3, so no structured RS framework using this noise on SST groups can certify more. HRS certifies ℓp\ell_p-bounded changes to at most rr entities. Deleting a group is an ℓ2\ell_2 change of 2.8–15.6 (Table 22), far above the ≈0.65 that ℓ2\ell_2 smoothing certifies. That is why the ℓ2\ell_2 route certifies 0 groups (Table 3).

The physical-loss certificates, no. Every structured RS method certifies a fixed set Ak(X)\mathcal{A}_k(X). EventCert certifies a probability over a random loss process: Pr⁡S∼Π[g(DS(X))≠A^]≤Pr⁡S∼Π[H>kwc(X)],H=∣S∩N(X)∣.\Pr_{S\sim\Pi}[g(\mathcal{D}_S(X))\neq\hat A]\le\Pr_{S\sim\Pi}[H > k_{\mathrm{wc}}(X)],\qquad H=|S\cap\mathcal{N}(X)|. This needs three ingredients that those frameworks do not have:

  1. The active-support identity DS(X)=DS∩N(X)(X)\mathcal{D}_S(X)=\mathcal{D}_{S\cap\mathcal{N}(X)}(X). It depends on the input and means that only event-carrying groups count (Sec. 4.3).
  2. Exact tails of HH. For uniform loss, this is the hypergeometric tail (Eq. 6). For a Gilbert burst channel, it is an O(mk)O(mk) forward recursion over the channel state and the number of lost active groups. A monotone coupling makes the certified loss rate well defined (Prop. 1, App. F.11).
  3. The hybrid worst case. When an adversary deletes aa groups before the channel acts, the worst choice is b⋆=min⁡(a,nX)b^\star=\min(a,n_X) active groups, proved by coupling (Prop. 2).

These certificates show that independent-loss guarantees overstate safety. On N-MNIST at 5% loss, certified accuracy is 85.6% under independent loss but only 14.5% under bursts (Table 5).

Table 3 already compares three RS baselines (randomized ablation on our groups, RS-Del-style deletion, Gaussian) with the same models and budgets. EventCert certifies 3.8–5.6× more. At pkeep=0.1p_{\mathrm{keep}}=0.1, ablation's ceiling is 6 groups and ours is 45.

Weakness 2 and Question 2: Smoothed accuracy vs. ACR across pkeepp_{\mathrm{keep}}.

From our existing logs (seeds 0–2, certified at every pkeepp_{\mathrm{keep}}). An asterisk marks the selected value, which for SHD is 0.2 or 0.3 depending on the seed.

p_keep          0.1     0.2     0.3     0.5     0.7     0.9
N-MNIST    acc  99.0*   99.3    99.4    99.6    99.6    99.6
           ACR  18.13*  15.20   11.39   5.82    2.95    1.97
DVS-Gest.  acc  69.6*   72.9    73.7    75.1    75.5    76.9
           ACR  7.28*   5.88    4.70    2.93    1.75    1.22
N-Caltech  acc  46.2*   51.6    53.4    54.7    54.5    54.3
           ACR  10.95*  6.73    4.81    2.54    1.36    0.93
SHD        acc  70.8    80.3*   82.7*   84.9    85.6    85.0
           ACR  2.05    2.49*   2.48*   2.08    1.52    1.14
N-TIDIG.   acc  44.4    65.1    71.0    75.7*   78.1    78.4
           ACR  0.59    0.79    0.93    1.12*   1.08    0.90
  • Vision: the trade-off is monotone, and the selected pkeep=0.1p_{\mathrm{keep}}=0.1 sits at the robust end. Compared with pkeep=0.9p_{\mathrm{keep}}=0.9, it gives up 0.6, 7.3 and 8.1 accuracy points for 9.2×, 6.0× and 11.8× the ACR.
  • Audio: ACR peaks at an intermediate value, because accuracy collapses under heavy masking (44.4% on N-TIDIGITS18 at 0.1). Selecting by ACR finds this knee automatically.
  • Accuracy first: on N-MNIST at pkeep=0.5p_{\mathrm{keep}}=0.5, EventCert reaches 99.6% accuracy with an ACR of 5.82. That is still above the best group-ablation ACR (4.45, Table 3).

We will add this table to Appendix D.

Summary

ConcernVerdictEvidence
Weakness 1: novelty relative to structured RSClarified. The worst-case layer is standard and optimal. The physical-loss certificates are new.l. 59–68, Eq. 3, 6–7, Props. 1–2
Question 1: can existing frameworks derive the certificates?Only the worst-case layer. Random, bursty and hybrid loss need new theory.Sec. 4.3–4.5, App. F.11
Weakness 1: no comparison with structured RSAlready in the paper. Three RS baselines. EventCert beats the best by 3.8–5.6×.Table 3
Weakness 2 and Question 2: accuracy–robustness trade-offAnswered with new data. The full table is above. The selected value is the ACR optimum.Table above, Fig. 1

We hope this resolves both questions, and we respectfully ask you to reconsider your score.

AI disclosure: The language has been polished and improved by AI. All analysis and computational work were done by the authors.

Thank you for your response and the additional results addressing my previous concerns. The newly provided results of accuracy–robustness trade-off resolve most of my concerns about the point. However, as I still have concerns about the novelty of the proposed method meets the high standards of ICLR, I will maintain my original score for now.

AuthorsOpenReview ↗

Clarification Requested on Remaining Novelty Concerns

I understand your decision. However, could you please clarify your remaining concerns regarding novelty? You mentioned the "high standards of ICLR." Could you specify which novelty criteria our work fails to meet?

As explained in our rebuttal, your previous novelty concerns have been directly addressed, and we believe the novelty of our contributions is technically justified. Novelty should be assessed based on concrete technical arguments and evidence, rather than a subjective perception of what constitutes a "high standard." We would appreciate specific technical reasons for maintaining this concern.

Reviewer_smZEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors proposed to certify the event-based classifiers under adversarial, burst, and combined attacks. The datasets used are the ones designed for the spiking neural nets processing.

Critical strengths
  • An approach to certify the loss of information in terms of pixels and time - and the authors claim that they are the first to do it
  • Taking into account different types of losses - in addition to spatio-temporal, adversary, random, bursty, and the combination was investigated
  • Seemingly correct theory involved
  • Reasonable and non-vacuous estimations are provided
  • Finding: clean training is not enough (the usual claim that has the roots since Cohen's work)
Critical weaknesses
  • To me, the approach seems like an extension to the mentioned papers of [1] and [2] where the authors used the combinatorial approach to certify either pixel-level robustness [2] or patch-level robustness [1]. Moreover, the main comparison - see the Table 3 - is again underlines it
  • In order to provide a clear separation of the method - other than the different set up (including groups and time) - to the mentioned [1] and [2], I think some reasoning is needed. So actually it seems that instead of structuring in the space domain like in [1], the authors of EventCert did the structuring in time domain.

[1] Alexander Levine and Soheil Feizi. (De)Randomized smoothing for certifiable defense against patch attacks. In Advances in Neural Information Processing Systems, volume 33, 2020.

[2] Alexander Levine and Soheil Feizi. Robustness certificates for sparse adversarial attacks by ran- domized ablation. In Proceedings of the 34th AAAI Conference on Artificial Intelligence, pp. 4585–4593, 2020.

Questions for authors

Please see the "Weaknesses" section. There is a concern that the paper is simply the extension of the previous work to the time domain.

1 comment
AuthorsOpenReview ↗

Clear separation from [1] and [2]: provable ceiling of 6 vs. 45, and certificates for physical loss

Thank you for the review and for rating the theory as correct. Your main concern is that EventCert extends [1] or [2] to the time domain. Four facts from the paper show that it does not.

1. Table 3 tests this directly, and the answer is no.

The row "Randomized ablation (Levine & Feizi, 2020b)" is [2], applied to our SST groups with the same architecture, training, tuning and Monte Carlo budget (l. 260–266). The two models are equally accurate. On N-MNIST at pkeep=0.1p_{\mathrm{keep}}=0.1, smoothed accuracy is 98.9% for [2] and 99.0% for EventCert. Even so, EventCert certifies 3.8–5.6× more groups on every dataset. The gap comes from the certificate itself, and it can be proved: [2]: pAL−pcompU>2(1−(m−ks)/(ms)),EventCert: pAL−pcompU>1−(1−pkeep)k.\text{[2]: } p_A^L-p_{\mathrm{comp}}^U > 2\Big(1-\tbinom{m-k}{s}\big/\tbinom{m}{s}\Big),\qquad \text{EventCert: } p_A^L-p_{\mathrm{comp}}^U > 1-(1-p_{\mathrm{keep}})^k . The left-hand side can never exceed 0.9915 at n=2048n=2048. So [2] can certify at most 6 groups at pkeep=0.1p_{\mathrm{keep}}=0.1, while EventCert can reach 45. The largest radii we measured on N-MNIST are exactly 6 and 45. An extension of [2] could not lift its ceiling 7.5-fold on the same units.

2. [1] does not fit this threat.

Derandomized smoothing certifies one contiguous patch. Our adversary can delete any kk groups (Sec. 3, l. 121–123). Suppose there are NN ablations, each seeing a fraction pp of the groups. Then kk scattered deletions can flip kpNkpN votes, so N≥nA−nB≥2kpN;⟹;k≤⌊1/(2p)⌋=5(p=0.1),N\ge n_A-n_B\ge 2kpN\\;\Longrightarrow\\;k\le\lfloor 1/(2p)\rfloor=5\quad(p=0.1), even for a perfect classifier. At the same keep rate, EventCert certifies a median of 17 groups and up to 45 on N-MNIST.

3. The structure is not only in time.

Each SST group spans time, space and polarity together (Sec. 3, l. 117–118). In Table 15, a group is (Δt,Δh,Δw)=(2,7,7)(\Delta t,\Delta h,\Delta w)=(2,7,7). EventCert certifies spatio-temporal blocks wherever they fall. It does not simply move the spatial structure of [1] into time.

4. Neither [1] nor [2] can express our contributions (l. 59–68).

Both certify a fixed threat set Ak(X)\mathcal{A}_k(X). EventCert also certifies random physical loss: Pr⁡S∼Π[g(DS(X))≠A^]≤Pr⁡S∼Π[,∣S∩N(X)∣>kwc(X),].\Pr_{S\sim\Pi}\big[g(\mathcal{D}_S(X))\neq\hat A\big]\le\Pr_{S\sim\Pi}\big[\\,|S\cap\mathcal{N}(X)| > k_{\mathrm{wc}}(X)\\,\big]. We compute this probability exactly in three settings:

  • Uniform loss (Eq. 6).
  • A Gilbert burst channel (Prop. 1), using an O(mk)O(mk) dynamic program over the channel state and the number of lost active groups.
  • An adversary acting together with a lossy channel (Prop. 2). The worst case, b⋆=min⁡(a,nX)b^\star=\min(a,n_X), is proved by coupling.

This gives a finding that neither [1] nor [2] could produce. On N-MNIST at a 5% loss rate, a guarantee that assumes independent loss reports 85.6%. Bursts of mean length 16 bring the true certified accuracy down to 14.5% (Table 5).

Summary

ConcernVerdictEvidence
Weakness 1: "an extension of randomized ablation [2]"Incorrect. [2] on our groups certifies 3.8–5.6× less. Its ceiling is 6 groups against our 45, at matched accuracy.Table 3, Sec. 5.3
Weakness 1: "an extension of derandomized smoothing [1]"Incorrect. [1] certifies one contiguous patch. For scattered deletions it is capped at 5 groups.Sec. 3, analysis above
Weakness 1: "Table 3 underlines the overlap"The opposite. Table 3 tests the overlap and rejects it.Table 3
Weakness 2 and Question: "structuring in the time domain"Incorrect. SST groups span time, space and polarity together.Sec. 3, Table 15
Weakness 2: separation from [1] and [2]Missed. Certificates for random, bursty and hybrid loss, which [1] and [2] cannot express.Eq. 6–7, Props. 1–2
Practical consequenceMissed. Guarantees that assume independent loss overstate safety, dropping from 85.6% to 14.5% under bursts.Table 5
Strengths: theory and estimatesAgreed. Correct and non-vacuous, as you note.App. F

At most, [1] and [2] give the worst-case layer, and their version of it is provably weaker. Since you found the theory correct and the estimates non-vacuous, we respectfully ask you to reconsider your score.

AI disclosure: The language has been polished and improved by AI. All analysis and computational work were done by the authors.

Reviewer_2iPaOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors study the problem of certification of streaming systems and describe a method to get upper bound of number of deletions in event tensor that is correct, given a predefined confidence level. Authors provide a way to get probabilistic guarantees on the model under losses which simulate real world situations.

Critical strengths
  1. The problem is well-motivated, the method seems correct to me.
  2. Extensive experimental evaluation, ablation study.
Critical weaknesses
  1. The method itself does not seem to be new, but a combination of existing well-known techniques (Bernoulli deletion, randomised ablation). The contribution is thus choosing the group and tail computations.
  2. Evaluation datasets are tiny, and the models are small; it is important to demonstrate whether the certification would yield comparable results on larger models.
  3. (I) seems to be known relevant work excerpted from the comparison; however, they appear suitable more that part of the baselines.

(I) Efficient Robustness Certificates for Discrete Data: Sparsity-Aware Randomized Smoothing for Graphs, Images and More by Aleksandar Bojchevski, Johannes Gasteiger, Stephan Günnemann (ICML 2020)

Questions for authors
  1. Run block/column derandomized smoothing and sparse smoothing (Bojchevski et al.) at group level as baselines. How are the results compare to their approach?
  2. Having n=2176n=2176 forward passes per input, how to use the certified models in a realtime?
  3. Please reframe the contribution: EventCert’s worst-case certificate resembles RS-Del &\& sparse-smoothing deletion certification with groups as tokens.
2 comments
AuthorsOpenReview ↗

Novelty clarified: certifying physical loss, which no RS framework provides

We thank the reviewer. Each concern is answered with evidence already in the paper.

Weakness 1/Question 3: Novelty vs. RS-Del / sparse smoothing

For the worst-case layer, we agree: the paper credits it (l. 45–48) and does not list it as a contribution (l. 59–68). It is also optimal. For Bernoulli group deletion the likelihood ratio is Pr⁡[DS(X)(M)=z]/Pr⁡[X(M)=z]∈0,qk−1\Pr[\mathcal{D}_S(X)^{(M)}=z]/\Pr[X^{(M)}=z]\in\\{0,q_k^{-1}\\} with qk=(1−pkeep)kq_k=(1-p_{\mathrm{keep}})^k, so Neyman–Pearson gives exactly Eq. 3. Sparse smoothing with p+=0p_+=0 on groups therefore returns identical radii, and no certificate with this noise can do better.

The contribution is what no RS framework (RS-Del, sparse, ablation, derandomized, HRS) provides: a guarantee against a random loss process, Pr⁡S∼Π[g(DS(X))≠A^]≤Pr⁡S∼Π[,∣S∩N(X)∣>kwc(X),].\Pr_{S\sim\Pi}[g(\mathcal{D}_S(X))\neq\hat A]\le\Pr_{S\sim\Pi}[\\,|S\cap\mathcal{N}(X)| > k_{\mathrm{wc}}(X)\\,]. It rests on three parts:

  • the active-support identity DS(X)=DS∩N(X)(X)\mathcal{D}_S(X)=\mathcal{D}_{S\cap\mathcal{N}(X)}(X),
  • exact tails for uniform loss (Eq. 6) and for Gilbert-burst loss (Prop. 1, via a DP with a monotone coupling),
  • the worst hybrid adversary, b⋆=min⁡(a,nX)b^\star=\min(a,n_X) (Prop. 2).

This yields the paper's key finding: at 5% loss, certified accuracy on N-MNIST falls from 85.6% to 14.5% when losses arrive in bursts (Table 5). We will state this explicitly and cite Bojchevski et al. (2020).

Question 1: Baselines

  • Sparse smoothing is identical to EventCert at group level. At entry level (p+=0p_+=0) it is our entry-masking row (Table 3, ACR ≤ 0.09), because one group holds more entries than that certificate covers (Table 22).
  • Derandomized smoothing: with NN ablations, each seeing a fraction pp of the groups, kk scattered deletions flip kpNkpN votes. Then N≥nA−nB≥2kpN⇒k≤⌊1/(2p)⌋=5N\ge n_A-n_B\ge 2kpN \Rightarrow k\le\lfloor 1/(2p)\rfloor=5 at p=0.1p=0.1, even for a perfect classifier. EventCert certifies up to 45 groups, with a median of 17 on N-MNIST.
  • Randomized ablation (Table 3) needs a margin of 2(1−(m−ks)/(ms))2(1-\binom{m-k}{s}/\binom{m}{s}), while ours needs only 1−qk1-q_k. With the margin at most 0.9915 (n=2048n=2048), its ceiling is 6 groups against our 45. The measured maxima are exactly 6 and 45, at matched accuracy (98.9% vs. 99.0%).

Weakness 2: Model size.

The certificate uses only class counts, and its ceiling depends only on (n,K,α,pkeep)(n,K,\alpha,p_{\mathrm{keep}}) (Table 16, App. F.9). A larger model can only raise accuracy under masking, and an SNN gives the same picture (Table 19). The evaluation used 4.6×1084.6\times10^8 forward passes on standard event benchmarks with up to 101 classes.

Question 2: Real time.

The 2176 passes run as one batch, taking 0.06–0.70 s per input on a shared laptop GPU (Table 24). That is 49× fewer samples than Cohen et al. (n=105n=10^5). Using n=512n=512 cuts the cost 4× and still certifies up to 32 groups (Fig. 2d). The physical-loss certificates need zero extra passes.

Summary

ConcernVerdictEvidence
"Bernoulli deletion + randomized ablation"Incorrect. Ablation on our groups certifies 3.8–5.6× less. Its ceiling is 6 vs. 45Table 3
Worst-case layer ≈ RS-Del / sparse smoothingAgreed, by design. Not claimed, provably optimall. 59–68, Eq. 3
Actual contributionMissed. Certificates for random, bursty and hybrid loss, 85.6% → 14.5% under burstsEq. 6–7, Props. 1–2, Table 5
Sparse-smoothing baselineAlready covered. Identical radii, the entry-level version is entry maskingTables 3, 22
Derandomized baselineNot applicable. Capped at 5 groupsAbove
Small modelsDoes not affect the certificate. Validity and ceiling are model-independent, holds for an SNNTables 16, 19
Real-time costAlready reported. 0.06–0.70 s/input, physical-loss certificates need no extra passesTable 24
Missing Bojchevski et al.Agreed. Will be cited—

You found the method correct and the evaluation extensive. That is why we respectfully ask you to reconsider your score.

AI disclosure: The language has been polished and improved by AI. All analysis and computational work were done by the authors.

Reviewer_2iPaOpenReview ↗

Response to authors

I thank authors for their response. However, I can not lean towards acceptance due to what the other revivers also mention – limited novelty of the approach.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary for reviewers and AC: shared novelty concern, the physical-loss contribution, and new trade-off data

To all reviewers and the AC

Thank you all for the careful reviews. Every reviewer rated soundness as "minor or fixable" and found the theory correct. All three raised the same concern, which is novelty relative to structured randomized smoothing. We answer it here, and the individual responses give the details.

  1. The worst-case layer is standard by design, and it is optimal. For Bernoulli group deletion, the Neyman–Pearson lemma gives exactly Eq. 3, so no certificate using this noise can certify more. We never list it as a contribution (l. 59–68).
  2. The contribution is certifying physical loss, which no RS framework does. Existing frameworks certify a fixed set Ak(X)\mathcal{A}_k(X). EventCert certifies Pr⁡S∼Π[g(DS(X))≠A^]≤δ\Pr_{S\sim\Pi}[g(\mathcal{D}_S(X))\neq\hat A]\le\delta for uniform, bursty and adversary-plus-channel loss (Eq. 6–7, Props. 1–2), without any extra forward passes. This leads to a new, safety-relevant finding. Guarantees that assume independent loss overstate robustness several-fold, falling from 85.6% to 14.5% on N-MNIST once losses arrive in bursts (Table 5).
  3. Randomized ablation [2] on our groups is already in Table 3, and it is provably weaker. With the same models and matched accuracy (98.9% vs. 99.0%), it can certify at most 6 groups at pkeep=0.1p_{\mathrm{keep}}=0.1, while EventCert reaches 45. The measured maxima are exactly 6 and 45. Derandomized patch smoothing is capped at 5 groups for scattered deletions.
  4. New data for Reviewer d7o2. We report smoothed accuracy against ACR for every pkeepp_{\mathrm{keep}} and dataset, taken from our existing logs.

Planned revision. We will reword the contribution statement so that it names the worst-case layer as the Bernoulli-deletion certificate applied to SST groups. We will also cite Bojchevski et al. (2020) and add the trade-off table and the ceiling analysis to the appendix.

AI disclosure: The language has been polished and improved by AI. All analysis and computational work were done by 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.