acceptodds
Under review as a conference paper at ICLR 2027

When Stronger Triggers Backfire: A High-Dimensional Theory of Backdoor Attacks

Abstract

Backdoor poisoning attacks behave counter-intuitively in high dimensions: when the test-time trigger budget is held fixed, a stronger training trigger can help the defender. We study generalised linear models on Gaussian-mixture data in the proportional regime (), varying the training trigger strength . Three phenomena emerge: (i) clean test accuracy can increase with ; (ii) the trigger alignment that drives attack success peaks at a finite and then declines; and (iii) low-variance directions can yield stronger trigger alignment. We derive closed-form asymptotic alignments for squared loss, including a unique finite peak for arbitrary mean-trigger overlap, and extend the finite-peak result and large-strength decay to a broader class of convex losses under suitable assumptions. A margin-variance decomposition identifies a finite-sample term that helps explain (i) and is absent from the population limit. Experiments on Gaussian surrogates and logistic regression on CIFAR-10 support the theoretical predictions, while ResNet-18 experiments exhibit similar qualitative behaviour beyond the convex setting.

Primary area:
learning theory
Keywords:
backdoor attacks, data poisoning, high-dimensional asymptotics, generalized linear models

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_1JwvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Despite the strong threats from data-poisoning backdoors, their behaviors are actually counter-intuitive with varying the trigger strength. Inspired by these, the author analyze the backdoor's behaviors and show that, with a fixed test trigger, stronger trigger strength can improve the clean accuracy and meanwhile reduce backdoor's attack success rate beyond an optimal strength. The proposed high-dimensional theory also identifies low-variance data directions as the most effective trigger directions, with experiments supporting these findings.

Critical strengths
  1. Insightful findings of backdoor attacks

The impact of the backdoor trigger strength on clean accuracy and attack success rate is very interesting. The increasing clean accuracy and a peak attack success rate also unveil a certain entanglement and disentanglement between the natural performance and the backdoor.

  1. The theoretical analysis on the three high-dimensionalities is rigorous.

The extensive analysis of the three impact factors of backdoor attacks provides a strong theoretical foundation for the considered attacking situation.

Critical weaknesses
  1. Unclear definition of trigger strength α\alpha

The trigger strength α\alpha depends on multiple factors, including the trigger opacity, the position on the input, the dissimilarity to the benign samples, the poisoning ratio, etc. Particularly in Section 3, the measurement of trigger strength α\alpha is unclear.

  1. Discussion of the theory for the application

In the abstract begin, the authors claim that the backdoor's counter-intuitive behaviors can help the defender. In the paper, however, there is no discussion about the application possibility of those interesting findings on defense methods.

  1. Limited consideration of attack variants.

Section D offers the poisoning setting, where a simplistic 2×22 \times 2 patch trigger is applied in this work. Given various existing poisoning methods, e.g., Blend [1] and WaNet [2], the adaptability of the proposed high-dimensional theory is not evaluated.

  1. Overly low clean accuracy of multiclass CIFAR10 with ResNet18.

In Figure 6, the clean accuracy with trigger norm from 0.000.00 to 1.501.50 remains lower than 90%. However, the standard training of a ResNet18 on CIFAR10 is normally at 94% ~ 95%. This lower clean accuracy requires further interpretation.

Questions for authors
  1. According to the visualization in Figure 1, the trigger strength with the highest attack success rate does not achieve a high enough clean accuracy. Can the authors explain what causes this mismatch?

  2. As shown in the evaluation, the logic regression models exhibits a clear alignment with the theoretical analysis in Section 3. In contrast, as mentioned by the authors, a ResNet18 exhibits similar qualitative behaviour beyond the convex setting. Can the authors interpret this problem from the viewpoint of the model capacity difference?

1 comment
AuthorsOpenReview ↗

Response to reviewer 1Jwv

We thank the reviewer for their review and are glad they found the findings insightful.

Trigger strength. α\alpha is the ℓ2\ell_2 norm of the additive perturbation αv\alpha v, with ∥v∥=1\Vert v\Vert =1 (Sec. 2, App. D). Position and shape are encoded in the direction vv, and the poisoning ratio is the separate parameter ϕ\phi. For our patch, vv is uniform on 12 entries (2×2×32\times2\times3). So α=0.5\alpha=0.5 adds ≈0.14\approx 0.14 and α=1.5\alpha=1.5 adds ≈0.43\approx 0.43 per entry, in [0,1][0,1] pixel units. We have restated this at the start of Sec. 3.

Defences. We agree with the reviewer that this deserves discussion. By "helps the defender" in the abstract, we mean that past the peak, clean accuracy is higher and ASR is lower; we now say this explicitly. We have also added a short paragraph on two implications:

  • The most effective training trigger has moderate strength (α⋆\alpha_\star), so a defender should not assume the most dangerous trigger is the most conspicuous one.
  • By Prop. 5, trigger alignment is controlled by gvv=v⊤(λI+τC)−1vg_{vv}=v^\top(\lambda I+\tau C)^{-1}v. For a linear squared-loss objective averaged over input noise, adding Gaussian noise of variance σn2\sigma_n^2 acts as extra ridge, capping gvv≤1/(λ+σn2)g_{vv}\le 1/(\lambda+\sigma_n^2).

Attack variants. A WaNet-style warp is in App. D (Fig. 7). Clean accuracy rises from 0.949 to 0.973, and ASR peaks at 0.76 near β=3.8\beta=3.8 then falls to 0.08. We now point to it from the introduction.

Multiclass accuracy. Our protocol is not standard CIFAR-10 training:

  • no data augmentation, because random crops and flips would move or remove the corner patch;
  • 50 epochs on an 80/20 split of the pooled data;
  • 5% of training labels flipped to class 0, which at α=0\alpha=0 is pure label noise.

We now state this in App. D.

Q1. This is the trade-off itself. For population squared loss in the eigenvector setting of Cor. 4, both metrics depend on α\alpha only through r=hv/hμr=h_v/h_\mu: clean accuracy falls as rr grows while ASR rises. In that model the ASR-optimal strength is exactly where clean accuracy is lowest. With finite samples the noise floor also affects the full margin variance (Sec. 4), but the tension remains.

Q2. Our theory concerns fixed-representation GLMs. ResNet learns its representation, so capacity and optimisation may affect both trigger features and sensitivity. Its qualitative agreement is empirical evidence beyond the theory, not a guarantee for every cross-entropy model with weight decay. The main difference we see between the two models is that the ResNet's ASR rises and falls much more sharply (Fig. 5).

We hope these clarifications address the reviewer's concerns.

Reviewer_4HULOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates an counterintuitive phenomenon in backdoor attacks: when the trigger strength used during deployment is fixed, the stronger the attacker embeds the trigger during training, the worse the attack performance becomes—normal accuracy of the model recovers, while attack success rate peaks at a moderate intensity and then steadily declines. Moreover, this optimal intensity can be predicted by a theoretical formula. The authors also find that the strongest attacks occur when triggers are embedded along directions of minimal data variation. These patterns are first rigorously proven and quantitatively characterized in simplified theoretical models, and then validated across various models—from simple logistic regression to deep networks (ResNet-18)—on the CIFAR-10 dataset, demonstrating that they are not artifacts of specific mathematical models but rather general principles with practical significance.

Critical strengths
  1. The phenomena observed in this paper are novel and counter-intuitive.

  2. The paper follows a well-organized narrative arc: it starts from empirical observations, explains them through a GLM-based theoretical analysis, and returns to deep networks for validation, forming a complete and closed-loop story.

  3. The authors demonstrate solid theoretical proficiency and carry out sound derivations.

Critical weaknesses
  1. The writing is hard to follow. For example, the contribution section introduces a large number of new concepts without explaining any of them nearby. The variables K and π in Eq. (1) are also left unexplained.

  2. The paper provides a comprehensive quantitative theoretical analysis of the three observed phenomena, but lacks a dedicated section that qualitatively and intuitively explains why these phenomena arise.

  3. The paper lacks a section summarizing the practical implications of its findings for deploying real-world backdoor attack and defense algorithms.

  4. The theory predicts how the peak position and peak magnitude depend on the poisoning ratio, the ratio of dimension samples, and the regularization strength, among other factors. However, most real data experiments use a fixed poisoning ratio. The research in Figure 3 regarding the dimension sample ratio only includes theoretical curves. Therefore, more parameter analyses with real experimental data are needed.

  5. On lines 458-462, it is concluded that the ASR will reach its peak at a certain finite value of α\alpha. However, the existence of a finite peak in hvh_v alone does not directly lead to this conclusion, as hμh_\mu and σ\sigma will also change with α\alpha. To rule out the possibility of mutual compensation among these quantities, or the occurrence of multiple peaks in ASR, additional conditions seem to be required.

  6. Font sizes should be made consistent across all figures. For instance, the font in Figure 2 is noticeably too large.

Questions for authors

N/A

2 comments
AuthorsOpenReview ↗

Response to reviewer 4HUL

We thank the reviewer for their positive assessment and helpful suggestions.

Clarity. In the revised paper we have:

  • defined KK and πc\pi_c after Eq. (1);
  • defined alignment, noise floor and information limit in one line each where they first appear;
  • harmonised figure fonts.

Intuition. We have added a short paragraph:

  • (i) A poisoned point needs positive margin. At large α\alpha, a weight of order 1/α1/\alpha on vv suffices, so the regulariser shrinks the vv-component and alignment falls.
  • (ii) Large-α\alpha poisoned points stop pulling θ\theta away from μ\mu, so θ\theta returns towards the clean solution. With finite nn, the noise floor ζ\zeta makes accuracy sensitive to this.
  • (iii) Weight on vv adds margin variance proportional to v⊤Cvv^\top Cv on every sample, so vv is cheapest to use when it is a low-variance direction.

Implications. Two practical points:

  • The most effective training trigger has moderate strength (the alignment optimum α⋆\alpha_\star is predicted by Eq. (5)), so a defender should not assume the most dangerous trigger is the most conspicuous one.
  • By Prop. 5, trigger alignment is controlled by gvv=v⊤(λI+τC)−1vg_{vv}=v^\top(\lambda I+\tau C)^{-1}v. For a linear squared-loss objective averaged over input noise, adding Gaussian noise of variance σn2\sigma_n^2 acts as extra ridge, capping gvv≤1/(λ+σn2)g_{vv}\le 1/(\lambda+\sigma_n^2).

Real-data sweeps. We have added a ResNet-18 sweep over ϕ∈{0.025,0.05,0.1,0.2}\phi\in\lbrace 0.025,0.05,0.1,0.2\rbrace on CIFAR-10 (App. D). At every ϕ\phi, clean accuracy rises with α\alpha and ASR peaks and then falls, and the peak moves to weaker triggers as ϕ\phi grows (from α≈0.55\alpha\approx0.55 to α≈0.24\alpha\approx0.24), qualitatively consistent with the linear-model trend.

ASR peak (l. 458–462). We agree with the reviewer that the alignment peak alone does not imply this. For squared loss in the eigenvector setting of Cor. 4, with sμ2>0s_\mu^2>0 and fixed αtest>0\alpha_{\text{test}}>0, new Cor. 9 proves a finite positive ASR maximiser in the proportional regime (App. A.6): ASR approaches the limiting classifier's clean error from above. For population squared loss with sμ2,sv2>0s_\mu^2,s_v^2>0, the peak is also unique at α2=(1−2ϕ)(λ+sv2)/(ϕ(1−ϕ))\alpha^2=(1-2\phi)(\lambda+s_v^2)/(\phi(1-\phi)). We do not prove a finite ASR maximum or its uniqueness for general convex losses.

We thank the reviewer again for their suggestions, which have improved the clarity of the paper.

Reviewer_4HULOpenReview ↗

Response

Thank you for the author's reply. This part has answered my confusion. Considering that the score I have given is already high enough and my confidence level is still insufficient, I will maintain this score.

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

The authors give theoretical analysis of the effect of the backdoor trigger norm ("strength") and backdoor trigger direction in high dimensions, considering the binary (2-class) case, with Gaussian classes, a linear classifier, and a convex loss objective. Given fixed test-time attack strength, the authors investigate the effect of varying the training-time attack strength (alpha). They (to me, unsurprisingly) find that the attack success rate peaks at a finite training attack strength. This is unsurprising, especially in the multi-class case, where increasing alpha may cause triggered patterns to ``collide'' with classes that are not the target class of the attack. What the authors do not discuss is that the maximum ASR appears to occur where alpha = alpha_test=0.5 -- their analysis results in a maximum alpha -- alpha_star -- but they do not discuss how close alpha_star is to alpha_test. The authors also show (unsurprisingly) that, to achieve high ASR, the trigger should be chosen not to align with principal signal directions. The authors also show (this is somewhat surprising) that clean generalization is maximized (approaches clean accuracy of an unpoisoned model) by sending alpha to infinity -- the explanation here is that for large alpha triggers are easily, correctly discriminated to the target class -- so less model capacity is wasted on them than when alpha is small/finite. However, it should also be noted that the loss in accuracy in Figure 1 is fairly small (less than 1%) for alpha=alpha_test.

The paper appears to be technically sound. My main concerns are: 1) the authors do not discuss how close alpha_star is to alpha_test -- empirically, it appears that the best training alpha equals alpha_test (which makes intuitive sense), but this is not discussed; 2) Unclear what the effect the poisoning rate has on these results (if this were varied); 3) Patch triggers (which model physical occlusion) would be more interesting than linear triggers -- but a patch mechanism is non-linear...the authors appear to use patch triggers in Figure 1, but they do not note that this is in fact a nonlinear trigger mechanism (not really consistent with their linear model).

Critical strengths

The authors perform sound theoretical analysis that provides insight into how clean generalization accuracy and ASR depend on the attack "strength" and how ASR also depends on the trigger direction.

The clean generalization accuracy phenomenon (approaching unpoisoned accuracy as alpha goes to infinity) is somewhat surprising.

Critical weaknesses

The authors do not discuss whether alpha = alpha_test is optimal, and how close alpha_star gets to alpha_test. Figures 1 and 2 are suggestive that the optimal alpha is alpha_test -- I would certainly expect this to be true in a multi (many) class setting, to avoid triggered samples from colliding with non-target classes...this seems very obvious, but it is not discussed by the authors... And from an operational standpoint it is simplest to choose alpha_test=alpha...

The model is quite simple (2-class, Gaussian classes, linear classifier, convex loss), but I understand that this is necessary to get results.

Questions for authors

What happens if one allows varying both alpha and the poisoning rate ? Will this change the optimal alpha ?

Is the trigger in Figure 1 additive (consistent) with the model or a patch trigger (nonlinear) ?

How can the authors' results inform a defense -- are there any existing detection defenses that "align" with the conclusions of this paper ?

3 comments
AuthorsOpenReview ↗

Response to reviewer pQWM

We thank the reviewer for their thoughtful review, and in particular for raising the relationship between α⋆\alpha_\star and αtest\alpha_{\text{test}}. We agree that a finite peak is intuitive once training and test strengths are separated.

α⋆\alpha_\star vs αtest\alpha_{\text{test}}. This is a fair point, and we should have discussed it. Matching α=αtest\alpha=\alpha_{\text{test}} is not optimal in general. In Fig. 5(a), logistic-regression ASR peaks at α≈0.55\alpha\approx 0.55 for all four values of αtest\alpha_{\text{test}}: at αtest=1\alpha_{\text{test}}=1, for example, ASR is 0.95 when training at 0.55 but 0.88 when training at α≈1\alpha\approx 1. The theory shows why matching need not be optimal. For population squared loss in the eigenvector setting of Cor. 4, ASR depends on α\alpha only through r=hv/hμr=h_v/h_\mu and increases with rr, so the ASR-optimal strength is α2=(1−2ϕ)(λ+sv2)/(ϕ(1−ϕ))\alpha^2=(1-2\phi)(\lambda+s_v^2)/(\phi(1-\phi)), independent of αtest\alpha_{\text{test}}. The closeness to 0.5 in Figs. 1–2 is specific to that setting. For ResNet-18 (Fig. 5b) the peak moves only from ≈0.3\approx 0.3 to ≈0.55\approx 0.55 while αtest\alpha_{\text{test}} grows from 0.2 to 1.25. We have added this to Sec. 4.

Effect size. We agree with the reviewer that the effect is modest: about 2 pp over the sweep in Fig. 1 and about 3 pp in Fig. 6. We claim it is systematic, not large.

Poisoning rate. Yes, it changes the optimum. By Eq. (5), α⋆2=(1+τgμμ)/(τϕgvv[1+τ(1−ϕ)gμμ])\alpha_\star^2=(1+\tau g_{\mu\mu})/(\tau\phi g_{vv}[1+\tau(1-\phi)g_{\mu\mu}]), with τ\tau independent of ϕ\phi (App. A.3), so the alignment peak moves to weaker triggers as ϕ\phi grows; the ASR optimum above does too. Fig. 4 and App. A.3 show this, and a new ResNet-18 sweep on CIFAR-10 confirms it on real data: the ASR peak moves from α≈0.55\alpha\approx0.55 at ϕ=0.025\phi=0.025 to α≈0.24\alpha\approx0.24 at ϕ=0.2\phi=0.2. For squared-loss ridge on CIFAR-10, Eq. (5) predicts the alignment peak at 0.68, 0.48 and 0.35 for ϕ=0.025\phi=0.025, 0.05 and 0.1, against observed peaks at 0.71, 0.47 and 0.32.

Trigger. It is additive, x↦x+αvx\mapsto x+\alpha v, with unit vv supported on the 2×22\times 2 corner (App. D). It is not a replacement or occlusion patch, and we now say "additive" in the Fig. 1 caption. A non-additive warp is in Fig. 7.

Defences. By Prop. 5, trigger alignment is controlled by gvv=v⊤(λI+τC)−1vg_{vv}=v^\top(\lambda I+\tau C)^{-1}v, which is largest along low-variance directions. For a linear squared-loss objective averaged over input noise, adding Gaussian noise of variance σn2\sigma_n^2 acts as extra ridge (Bishop, 1995), capping gvv≤1/(λ+σn2)g_{vv}\le 1/(\lambda+\sigma_n^2) in every direction. We have added this remark to the paper.

We hope these clarifications address the reviewer's concerns.

The authors have well-responded to my comments and questions. I have increased my score.

Regarding the following answer to the area chair: "A fixed deployed budget models an attacker whose training-time data manipulation can be stronger than the perturbation permitted at deployment." -- This answer is not really giving a practical scenario where there is an attack-strength "budget" -- it is much more likely, in practice, that there is a budget on the poisoning rate, not on the attack "strength"... the attack "strength" (both for poisoned training samples and for operational triggers) might be chosen relatively small in order to evade a detection defense...

Thank you for the positive assessment and this clarification. We agree that poisoning-rate constraints are important and that, as you note, evading detection can motivate small triggers during both training and deployment. Our question is whether matching the training strength to the deployed strength is optimal. We will clarify that our analysis holds with the deployed strength fixed and that the feasible range of training strengths depends on the threat model.

Reviewer_4nmaOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the effect of training-trigger strength on backdoor poisoning when the test-time trigger strength is fixed. Its central observation is that increasing the training trigger can initially improve an attack but may subsequently reduce its effectiveness. The theoretical model considers regularized generalized linear classifiers trained on a binary Gaussian mixture in the proportional high-dimensional regime. A fraction of negative-class training examples is modified by adding a trigger and flipping their labels to the target class. For squared loss, the paper derives asymptotic expressions for alignment with the benign mean and trigger directions. It establishes a finite peak in trigger alignment under the stated conditions and analyzes the effect of trigger direction through the covariance spectrum. For a broader class of decreasing convex losses, it studies the existence of a finite alignment maximum and the decay of alignment at large training-trigger strength. The empirical evaluation includes Gaussian surrogate data, logistic regression on CIFAR-10, and ResNet-18 experiments.

Critical strengths
  1. Separating training-trigger strength from a fixed test-time trigger budget reveals a meaningful attack-design trade-off. The possibility that an overly strong training trigger can reduce effectiveness under a constrained deployment trigger is relevant to understanding backdoor behavior.
  2. The margin-variance decomposition separates signal-related components from a fluctuation term that can remain non-negligible in the proportional regime. This is a useful lens for understanding why a population approximation may not reproduce the same clean-accuracy behavior.
  3. This paper specifies the poisoning model, trigger orthogonality conditions, and order of limits. It also acknowledges that the theory is developed for binary Gaussian-mixture GLMs and does not directly cover non-convex deep networks or general input-dependent triggers.
Critical weaknesses
  1. The paper argues that the high-dimensional noise floor makes accuracy sensitive to changes in benign alignment. This is plausible, and the reported variance decomposition is helpful. However, the displayed monotonicity calculation neglects the trigger component and does not by itself control how the noise floor and other variance terms vary with. In addition, this paper acknowledges that strict monotonicity of benign alignment for general-loss ERM remains open. The theoretical contribution is therefore better described as identifying a possible mechanism and proving selected alignment or endpoint results, rather than proving a general monotonic increase in clean accuracy.
  2. Proposition 5 maximizes leading-order trigger alignment among directions satisfying the stated orthogonality assumption. It does not establish that every minimum-eigenvalue direction is feasible, nor that the direction maximizing trigger alignment necessarily maximizes attack success after accounting for benign alignment and margin variance. Therefore, a direct comparison of alignment and ASR optima under the same assumptions would clarify the gap.
  3. The central phenomenon is studied by varying the training trigger while holding the test trigger fixed. This is a legitimate and clearly stated setting, and Appendix Figure 5 explores multiple fixed budgets. However, the conclusion should not be read as saying that stronger training triggers generally weaken attacks when the deployed trigger can grow with them. The practical relevance of the mismatch between training and test strengths would be stronger with examples of concrete deployment constraints and with matched-budget or adaptive-test-trigger comparisons discussed alongside the fixed-budget results.
Questions for authors
  1. In the clean-accuracy analysis, how much of the observed improvement remains when the trigger and cross terms are retained and the (\alpha)-dependence of the fluctuation term is included? Can this be shown over the full strength sweep rather than at a selected setting?
  2. Proposition 6 proves existence of a finite alignment maximum under the stated assumptions. Is this maximum unique for general convex losses, or can multiple local or global maxima occur?
  3. How sensitive are the main empirical trends to poisoning fraction, regularization strength, class pair, and trigger placement? The current main experiments focus primarily on a single poisoning fraction and one CIFAR-10 setting.
  4. How do the conclusions change when the attacker can adapt the deployed trigger within a realistic budget?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 4nma

We thank the reviewer for their careful reading and constructive comments.

W1 (clean accuracy). We agree with the reviewer. What we prove is:

  • leading-order monotone benign alignment for squared loss (Cor. 3);
  • for general losses, that benign alignment at α=0\alpha=0 lies strictly below its α→∞\alpha\to\infty limit in the population eigenvector setting (Prop. 7); Prop. 8 is a local one-step result.

We identify the mechanism for this in Sec. 4: the finite-sample noise floor ζ\zeta couples clean accuracy to hμh_\mu. Since ζ\zeta and the trigger term also vary with α\alpha, this does not rigorously prove that clean accuracy is monotone in α\alpha, and we have reworded Secs. 3–4 and the conclusion to say so explicitly.

On Q1: we solved the fixed-point equations of Theorem 1 over the full 20-point sweep with CIFAR-10 statistics, keeping every variance term and the α\alpha-dependence of ζ\zeta. Clean accuracy dips from 82.5% to 82.2% at α≈0.32\alpha\approx 0.32, then rises to 83.1% at α=1.5\alpha=1.5. The trigger and cross terms reach about 7% of the variance mid-sweep, so the recovery does not rely on dropping them. We have added this sweep to Fig. 2 and App. C.

W2 (Prop. 5). The reviewer is correct. It is an alignment result, and it applies only when a minimum-eigenvalue direction satisfies Assumption 2. ASR also depends on hμh_\mu and σ\sigma, and both vary with sv2s_v^2. Evaluating the ASR formula of Sec. 4 with Cor. 4, we find numerically (Fig. 8 setting and a random parameter sweep) that whenever the attack succeeds on at least half of triggered inputs (ASR ≥1/2\geq 1/2), lower-variance trigger directions give strictly higher ASR. Below 1/21/2, where the attack fails on most triggered inputs, the ordering can reverse, since extra margin variance then pushes ASR towards 1/21/2. We have added this ASR panel to Fig. 8 and stated it. The right panels of Figs. 1 and 6 measure ASR directly. The same gap affects the peak in α\alpha. There we have now added a new corollary (Cor. 9) proving that, for squared loss in the setting of Cor. 4, ASR itself has a finite maximiser for every deployed strength, and a second (Cor. A.1) proving that this maximiser is unique in the information limit.

W3 / Q4 (fixed budget). We agree with the reviewer that the backfire requires a constrained deployed trigger. Examples:

  • the trigger must stay small enough to be imperceptible, or to pass an input filter;
  • it is a physical trigger whose contrast the attacker cannot raise at test time.

For scale, our αtest=0.5\alpha_{\text{test}}=0.5 equals the CIFAR-10 ℓ2\ell_2 budget of 0.5 used by RobustBench (Croce et al., 2021), and an ℓ∞\ell_\infty cap of 8/2558/255 on our 12-entry patch gives αtest≈0.11\alpha_{\text{test}}\approx 0.11. We have added these budgets to Sec. 3. With αtest=α\alpha_{\text{test}}=\alpha (dashed line, Fig. 5), ASR increases with α\alpha in both models over the tested range. We have stated this scope in the abstract and introduction.

On Q4: in the Gaussian model, if the attacker can choose any deployed strength 0≤t≤T0\le t\le T along the same direction, ASR =Φ((thv−hμ)/σ)=\Phi((t h_v-h_\mu)/\sigma) increases with tt since hv>0h_v>0, so t=Tt=T is optimal and the fixed-budget curves already cover this case. Adapting the trigger direction would need a separate analysis.

Q2. Uniqueness of the peak for general losses is open. We prove it for squared loss, and the logistic curves have a single peak in every setting we solved (Figs. 2–3). We now say so after Prop. 6.

Q3. We have added a ResNet-18 sweep over ϕ∈{0.025,0.05,0.1,0.2}\phi\in\lbrace 0.025,0.05,0.1,0.2\rbrace on CIFAR-10 (App. D). At every ϕ\phi, clean accuracy rises with α\alpha and ASR peaks and then falls, with the peak moving to weaker triggers as ϕ\phi grows (from α≈0.55\alpha\approx0.55 to α≈0.24\alpha\approx0.24), qualitatively consistent with the linear-model trend. For the other factors, we ran squared-loss ridge on CIFAR-10 varying λ∈{10−5,10−4,10−3}\lambda\in\lbrace 10^{-5},10^{-4},10^{-3}\rbrace, the class pair (bird/cat) and the patch location (bottom-right). Every setting keeps an interior ASR peak, and the alignment-peak prediction from Eq. (5), evaluated from the data statistics with no fitted parameters, lies within one grid step of the observed ASR peak in settings with computed predictions (e.g. 0.60 vs 0.63 at λ=10−3\lambda=10^{-3}, 0.39 vs 0.39 for the bottom-right patch). The multiclass experiment (Fig. 6) and the 20 trigger directions of Figs. 1 and 6 (right) cover further class structure and trigger directions.

We hope these clarifications address the reviewer's concerns.

Reference. Croce et al. (2021). RobustBench: a standardized adversarial robustness benchmark. NeurIPS Datasets and Benchmarks Track.

I appreciate the authors' detailed response and the additional experimental results and theoretical analyses. The new full-range parameter sweeps, ASR analyses, and evaluations across different experimental settings address several of my previous concerns. I also acknowledge the substantial improvements made during the rebuttal.

However, I still have concerns that some key conclusions rely on specific model assumptions and constraints on the fixed test-time trigger strength. For example, the newly added full-range parameter sweep shows that clean accuracy is not always monotonic, while the optimality and uniqueness of the ASR maximum under general convex losses remain theoretically unresolved.

I acknowledge the substantial improvements made during the rebuttal. Although there still need more explain for improvement regarding theoretical generalizability and practical applicability. So I raise my rating from 2 to 3.

We thank the reviewer for updating their score and for their comments on the improvements during the rebuttal. We agree that a theoretical result in the non-convex case would be an interesting direction for future work. Although we have now shown finite optimality of the ASR maximum over the square loss, the general convex case and uniqueness outside the information limit remain open.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.