acceptodds
Under review as a conference paper at ICLR 2027

SCAR: Class-Wise Assessment and Shortcut-Guided Model Revision for Backdoor Defense

Abstract

Backdoor attacks pose a severe security threat to deep neural networks (DNNs) by implanting hidden behaviors activated by specific triggers. This threat is particularly concerning for downstream users acquiring models from untrusted sources, for whom reliable deployment entails assessing model integrity and, if compromised, localizing suspicious target classes and repairing the model. Backdoor learning introduces an additional trigger-to-target association beyond the intended task mapping, forming a shortcut that can leave class-specific irregularities in classifier-head weights and prediction responses, as well as directional changes in feature space. Building on these observations, we propose Shortcut-aware Class-wise Assessment and Revision (SCAR), a unified post-training framework for backdoor assessment and model recovery. SCAR first combines classifier-weight anomalies with perturbation-induced prediction preferences to assess model integrity and localize suspicious target classes. For each localized class, it then estimates a shortcut-related feature direction from class-specific weights and feature-wise variability, and selectively revises the corresponding classifier-head parameters along this direction to suppress backdoor behavior while preserving normal predictive utility. Extensive experiments across diverse datasets, attacks, and architectures demonstrate the effectiveness and generality of SCAR for both backdoor assessment and model recovery. Code will be released upon acceptance.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI Security, Backdoor Defense, Backdoor Detection, Model Repair

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_HWRTOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SCAR for inspecting and repairing backdoors in pretrained classification models obtained from untrusted sources. Given white-box access and a small unlabeled downstream dataset that may contain poisoned samples, SCAR combines classifier-weight anomalies with transformation-induced prediction preferences to identify suspicious target classes. It then estimates shortcut-related directions using filtered feature statistics and class-specific weights, refines these directions using class geometry, and directly revises the corresponding classifier-head parameters. Its core contribution is to connect class-wise backdoor assessment with targeted repair, without requiring the original training data, ground-truth labels, triggers, or additional training.

Critical strengths
  1. The model inspection and repair approach is novel and interesting. SCAR connects class-wise anomaly detection with targeted classifier-head revision without requiring triggers, ground-truth labels, or additional training, offering practical value for downstream users with limited resources.

  2. The experiments cover a broad range of models, datasets, and baselines. The evaluation includes four datasets, four architectures, ten attacks, and multiple baselines, and examines multi-target attacks, coexisting heterogeneous backdoors, and different downstream contamination rates, providing broad empirical support for the method's effectiveness.

Critical weaknesses
  1. The proxy estimation of the backdoor-related direction is insufficiently justified. The optimal direction in Eq. (11) assumes access to the true trigger-induced shift, whereas Eq. (12) constructs a proxy from dimensions with large positive class-specific weights and low feature variance, without explaining why the two should align. Such dimensions may encode benign semantics, and a backdoor may also operate by reducing activations in dimensions with negative weights. Figures 3 and 8 provide qualitative support, but optimality in the oracle setting does not directly establish the reliability of the practical direction estimate.

  2. The link between direction estimation and effective repair lacks a decision-level justification. Even if the direction captures trigger-induced changes, the correction in Eq. (14) must be large enough to overcome the backdoor margin while preserving benign classification margins; the paper does not establish when both requirements can be met. Centering at the mean class projection only ensures a zero average correction over the reference samples, and the geometric refinement does not ensure that alignment with the trigger-induced shift is preserved. Table 6 supports the average effectiveness of the combined direction, but the effectiveness of a feature direction alone does not establish the reliability of the classifier-head revision.

Questions for authors
  1. What is the specific experimental setup for Figure 1, and how are the quantities on the horizontal and vertical axes defined?

  2. How is the filtering threshold τf\tau_f in line 306 selected, and how sensitive are the results to it?

  3. Why does SCAR have relatively high FPR95 values for LC and SSBA on CIFAR-10 in Table 1?

  4. Do the baselines also use unlabeled defense data containing poisoned samples? How are baselines that require labels handled?

8 comments
AuthorsOpenReview ↗

Author Rebuttal Part (1/3)

We appreciate the reviewer's careful evaluation and insightful comments. We address the concerns and questions in detail below.

(1) Justification and reliability of shortcut direction estimation.

Motivation of the shortcut proxy. We first clarify the rationale for using class-specific weights and feature variability to estimate shortcut-related directions. As discussed in Sec. 4.2, backdoor learning introduces a trigger-to-target shortcut that strongly promotes target-class predictions when activated while remaining largely dormant on benign inputs. These two properties motivate our proxy design in Eq. (12). First, dimensions with large positive class-specific weights can efficiently increase the target-class response when activated, making them potential carriers of the backdoor shortcut. Second, the shortcut's limited influence on benign predictions motivates favoring dimensions with low feature variability, whose relatively stable responses under benign inputs may allow trigger-induced changes to affect target-class predictions without substantially disrupting normal classification. Although the selected dimensions may also encode benign semantics, our objective does not require a strict separation of benign and backdoor features, but rather a direction that captures trigger-induced changes relative to normal feature variability, as formulated in Eq. (9).

Reliability of practical direction estimation. We further assess the reliability of the estimated shortcut-related directions through oracle alignment and empirical defense performance. Figures 3 and 8 demonstrate positive alignment between the practical estimate scs_c and its oracle counterpart scoras_c^{\mathrm{ora}} across representative attacks and model architectures. Furthermore, extensive experiments across four datasets, four architectures, and ten attacks, together with additional evaluations against adaptive attacks (Reviewer aqd7, Q5), demonstrate effective backdoor mitigation with largely preserved clean accuracy, further supporting the reliability of the proxy-based direction estimation.


(2) Negative class-specific weights in shortcut direction estimation.

To examine the contribution of negative class-specific weights to backdoor-induced target-class predictions, we analyze ten attacks on CIFAR-10/ResNet-18 (C10) and CIFAR-100/VGG-16 (C100) using paired benign and triggered inputs. We consider two pathways through which trigger-induced feature changes increase the target-class relative logit: increasing activations in dimensions with positive class-specific weights and decreasing activations in dimensions with negative weights. Table 1 reports the proportion attributable to the latter among the positive logit contributions from these two pathways. Across all evaluated attacks, this proportion ranges from 2.30% to 20.09%, indicating that the former pathway consistently dominates. These findings provide empirical support for prioritizing positive-weight dimensions in Eq. (12).

Table 1. Percentage of trigger-induced target-class relative logit gains attributable to negative-weight dimensions.

AttackC10C100
BadNets15.2114.41
Blended18.9215.28
TrojanNN17.4416.37
LC16.712.30
SIG4.3318.96
SSBA19.4518.29
WaNet20.0912.87
BPP17.7719.59
LOTUS12.6311.42
Grond19.2414.05

To further examine the role of negative class-specific weights in shortcut direction estimation, we compare three variants using positive weights only (SCAR), negative weights only (SCAR−), and both (SCAR±). As shown in Table 2, SCAR achieves the lowest average ASR on both benchmarks while maintaining competitive benign ACC. Using only negative weights substantially weakens backdoor suppression, particularly on C10, while incorporating both signs does not further reduce ASR. These results support the effectiveness of the original design for backdoor mitigation. Nevertheless, classifier weights alone cannot determine the direction of trigger-induced feature changes, and simply incorporating negative weights does not necessarily improve the shortcut estimate. Exploring how to better leverage negative-weight information remains a promising direction for future research. We will include these experiments and further discussion in the revised paper and release our code to facilitate future work.

Table 2. Defense performance of SCAR variants using different weight signs, averaged over ten attacks.

MethodACC ↑ASR ↓ (C10)ACC ↑ASR ↓ (C100)
SCAR93.631.6972.030.93
SCAR−93.2987.0372.2514.54
SCAR±92.7610.5572.002.02
AuthorsOpenReview ↗

Author Rebuttal Part (2/3)

(3) Decision-level justification of classifier-head revision.

Decision-level conditions for effective revision. We agree that identifying a shortcut-related direction alone does not establish the effectiveness of classifier-head revision. To clarify this connection, we rewrite Eq. (14) by defining dc(x)=sc⋆⊤h(x)−qcd_c(x)=s_c^{\star\top}h(x)-q_c as the centered directional response and Δc(x)=γc⋆dc(x)\Delta_c(x)=\gamma_c^\star d_c(x) as the resulting logit correction. Since only the target-class logit is modified, the revised margin satisfies Mc⋆(x)=Mc(x)−Δc(x)M_c^\star(x)=M_c(x)-\Delta_c(x), where Mc(x)=fc(x)−max⁡i≠cfi(x)M_c(x)=f_c(x)-\max_{i\neq c}f_i(x). For triggered inputs predicted as class cc, successful suppression requires Δc(x)>Mc(x)\Delta_c(x)>M_c(x) to reverse their target-class preference. Conversely, preserving benign predictions of class cc requires Δc(x)<Mc(x)\Delta_c(x)<M_c(x). For benign inputs predicted as other classes, a positive correction further reduces the target-class logit without weakening their decision margins against class cc. Thus, effective revision depends not only on directional alignment but also on whether the induced corrections sufficiently distinguish triggered inputs from benign ones.

Role and limitations of reference centering. We acknowledge that centering at qcq_c guarantees only a zero-mean correction over the reference samples predicted as class cc, rather than individual margin preservation for arbitrary or unseen samples. SCAR does not rely solely on centering to preserve benign predictions. Instead, its direction estimation prioritizes feature dimensions with strong class-specific influence but relatively low variability across downstream samples, exploiting the observation that backdoor shortcuts are largely dormant on benign inputs. The subsequent geometric refinement incorporates the normal class structure to further reduce interference with benign predictions. These mechanisms collectively encourage a larger correction for triggered inputs than for benign ones, although they do not provide a formal guarantee of individual margin preservation.

Preservation of shortcut alignment after geometric refinement. We further investigate whether incorporating the class-geometry direction compromises the suppression capability of the shortcut-related direction. As detailed in our response to Reviewer KKQ3, the final directions retain positive alignment with both the oracle shortcut directions and trigger-induced feature shifts, even under relatively strong geometric correction (β=2\beta=2 and 88). Moreover, Table 6 of the main paper shows that the combined direction achieves an ASR of 0.93%0.93\%, compared with 1.68%1.68\% using the shortcut-related direction alone, while improving ACC from 70.33%70.33\% to 72.03%72.03\%. This suggests that geometric refinement can improve benign accuracy while largely retaining the shortcut-suppression effect, supporting a balance between backdoor mitigation and benign prediction preservation.

Direct evidence from classifier-head revision. More importantly, we provide direct decision-level evidence beyond the directional analysis in Table 6 of the main paper. As reported in our response to Reviewer aqd7 (Q4), the average target-class logit margin of triggered samples decreases from 9.079.07 to −14.16-14.16 after revision, indicating that the correction reverses their preference for the target class. Meanwhile, benign target-class and non-target-class samples retain positive and negative target-class margins, respectively, indicating that their original classification decisions are largely preserved. These results directly demonstrate that the estimated direction, together with the correction in Eq. (14), translates shortcut-related feature responses into effective and selective classifier-head revision. We acknowledge that these empirical results do not establish a universal guarantee for every individual sample, but they provide direct evidence of the decision-level effectiveness of SCAR under the evaluated settings.

Overall justification. Taken together, the decision-level analysis and empirical results provide complementary evidence for the effectiveness of SCAR. The estimated direction identifies feature responses associated with the backdoor shortcut, while the revision in Eq. (14) translates these responses into a targeted adjustment of the classifier's decision boundary. Although individual margin preservation cannot be guaranteed for arbitrary inputs, the results demonstrate that SCAR can substantially suppress triggered target-class predictions while largely maintaining benign classification behavior under the evaluated settings.

AuthorsOpenReview ↗

Author Rebuttal Part (3/3)

(4) Experimental setup and axis definitions of Figure 1.

Both panels in Figure 1 are obtained from a ResNet-18 trained for 200 epochs on CIFAR-10 under the Blended attack. Left: Each point represents a class, with the vertical axis showing the L2L_2 norm of its classifier weight and the horizontal axis indicating the percentage of benign test images from other classes classified into that class after mild Gaussian noise perturbation. Right: The horizontal axis indexes the first 32 dimensions of the estimated direction scs_c, while the vertical axis shows the feature responses of triggered samples and their benign counterparts at these dimensions. Solid lines represent the mean responses across samples, and shaded regions indicate their variance. We will clarify these details in the revised paper.


(5) Filtering threshold selection and sensitivity.

We determine the filtering threshold τf\tau_f using an attack-agnostic, data-driven strategy based solely on the acquired model and unlabeled downstream data. Specifically, we first exclude samples predicted as suspicious target classes, then compute the preference score of each remaining sample with respect to its predicted class and set τf\tau_f to the empirical 0.93 quantile of these scores. For example, on CIFAR-10/ResNet-18, this procedure yields similar thresholds across different attacks (mean: 29.72; range: 25.52–33.94), supporting the use of a shared threshold τf=30\tau_f=30 without attack-specific tuning.

To assess sensitivity to the filtering threshold, we vary the quantile from 0.90 to 0.96 under the BadNets attack on CIFAR-10/ResNet-18. As shown in Table 3, SCAR maintains relatively stable clean accuracy and effective backdoor mitigation across these settings, with the 0.93 quantile achieving the lowest ASR while preserving competitive ACC. These results suggest that SCAR is not overly sensitive to the threshold choice. We will clarify the threshold selection strategy in the implementation details and include the corresponding sensitivity analysis in Sec. 5.3 of the revised manuscript.

Table 3. Sensitivity analysis of the filtering threshold under BadNets on CIFAR-10/ResNet-18.

QuantileACC ↑ASR ↓
0.9094.644.00
0.9194.271.19
0.9394.500.13
0.9494.351.02
0.9693.713.17

(6) Analysis of elevated FPR95 under LC and SSBA attacks.

The elevated FPR95 under LC and SSBA on CIFAR-10/ResNet-18 may reflect both attack-specific characteristics and the dataset–architecture configuration. LC uses adversarially perturbed, label-consistent poisons, potentially making backdoor-related anomalies less distinguishable from benign class-specific behavior. SSBA employs imperceptible, sample-specific triggers, which may produce less consistent transformation-induced prediction preferences. These factors could increase the overlap between clean and backdoored model detection scores at high recall. Nevertheless, SCAR maintains high AUROC on CIFAR-10/ResNet-18, indicating strong overall discrimination between clean and backdoored models, while achieving perfect detection performance for both attacks on CIFAR-100/VGG-16. We will include this analysis in Sec. 5.2 of the revised paper.


(7) Data assumptions and label handling for defense baselines.

Table 4 summarizes the requirements for trusted clean data and ground-truth labels in the complete defense procedures of the baseline methods. Unlike SCAR, which operates on unlabeled downstream data that may contain poisoned samples, all baselines assume access to trusted clean data. In our experiments, we follow their original defense protocols, providing clean defense data and ground-truth labels where required, rather than substituting pseudo-labels or potentially contaminated data. Notably, some individual stages are label-free, including FP's pruning, NC's trigger inversion, and MM-BM's activation-bound optimization, although MM-BM's complete procedure uses labels for sample selection and accuracy constraints. We will include these clarifications in Appendix A.3 of the revised manuscript.

Table 4. Data and label requirements of defense baselines and SCAR.

MethodGround-truth labelsTrusted clean data
FP✓✓
NC✓✓
NAD✓✓
FT-SAM✓✓
RNP✓✓
MM-BM✓✓
TSBD✓✓
D3✓✓
PGBD✓✓
LMR✓✓
SCAR✗✗
Reviewer_KKQ3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SCAR connects post-training backdoor inspection with targeted classifier-head repair using a small, potentially contaminated unlabeled dataset. It combines weight anomalies and transformation-induced class preferences, then estimates a shortcut-related direction to revise only suspicious head parameters without retraining. Across four main dataset–architecture settings and ten attacks, the authors report average ASR below 2% with average clean-accuracy losses of 0.18–0.74 percentage points. Component, multi-target, and contamination analyses support the usefulness of the integrated workflow. The reported results establish strong mitigation within the tested data-poisoning scope; repair of direct weight-poisoning backdoors remains limited.

Critical strengths
  • Section 3 addresses a practical problem: a downstream user must assess and repair an untrusted model using limited, potentially contaminated unlabeled data. The interesting contribution is translating class-wise inspection evidence into targeted head revision without retraining. Inspection and repair are established tasks, but this lightweight connection is useful under the stated access constraints.

  • Tables 1–6 offer broad empirical coverage: ten attacks, four main dataset–architecture settings, component analyses, and multi-target cases. Baselines include SPD and LMR (2026), so the comparison is not confined to older defenses. The reported repair–accuracy trade-off supports practical value; the direction analyses do not, however, settle the proxy and geometry assumptions discussed below.

Critical weaknesses

The shortcut proxy and geometry correction in Equations (12)–(13) rely on alignment with trigger-induced feature shifts. Figures 3 and 8 support the estimated direction, but do not similarly examine the final corrected direction, particularly when geometry receives a large weight. This limits the mechanistic interpretation without negating the reported mitigation results.

Questions for authors

Can the existing oracle analysis confirm that the final corrected direction retains suppressive alignment with trigger-induced shifts, especially at beta = 8 and 2?

3 comments
AuthorsOpenReview ↗

Author Rebuttal

We appreciate the reviewer's insightful comments on the alignment of the final revision direction and address this concern below.

Directional alignment after geometry correction.

We extend the oracle analysis in Figs. 3 and 8 to examine whether the final revision direction retains suppressive alignment after geometry correction. Specifically, we compare the original direction scs_c (β=0\beta=0) with the geometry-corrected direction sc⋆s_c^\star (β=2,8\beta=2, 8) under LC attacks on ImageNet-1k/ResNet-50 and TinyImageNet/ViT-B/16. We report their cosine similarities with the oracle direction scoras_c^{\mathrm{ora}} and the trigger-induced feature shift Δμcora\Delta\mu_c^{\mathrm{ora}}, alongside ASR and ACC.

As shown in Table 1, geometry correction reduces but preserves positive alignment with both the oracle directions and trigger-induced shifts, indicating that the final revision still counteracts trigger-induced target-class responses on average. Notably, even the oracle direction is not intended to directly follow the raw trigger-induced shift, as it also accounts for normal task variation through the reference feature covariance. By further incorporating class geometry, SCAR improves benign accuracy by 1.49 percentage points on both benchmarks while maintaining low ASR at the adopted correction strengths (β=2\beta=2 for ImageNet-1k and β=8\beta=8 for TinyImageNet). However, further increasing β\beta on ImageNet-1k weakens directional alignment and raises ASR with negligible additional accuracy gains, suggesting that excessive correction may compromise shortcut suppression. Overall, these results support the adopted correction strengths as a balance between preserving benign accuracy and retaining effective backdoor suppression. We will include this analysis in Sec. 5.3 of the revised manuscript.

Table 1. Directional alignment and defense performance under different geometry correction strengths.

Dataset / Modelβ\betaCos. w/ OracleCos. w/ ShiftASR ↓ACC ↑
ImageNet-1k / ResNet-5000.820.672.5975.39
20.710.504.3076.88
80.460.398.0876.90
TinyImageNet / ViT-B/1600.720.640.0085.23
20.670.580.0185.88
80.590.420.0186.72

Thank you for the additional analysis. This directly answers my question. I choose to keep the rating at 3.

Thank you for your feedback and thoughtful consideration. We sincerely appreciate your time and effort in reviewing our work.

Reviewer_aqd7OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a unified framework for backdoor detection and mitigation, termed SCAR (Shortcut-aware Class-wise Assessment and Revision). It addresses the limitations of existing post-training backdoor defense methods, which often rely on trusted clean data, trigger reconstruction, or additional model training, and frequently lack an effective link between the detection and mitigation phases. The framework is designed for scenarios where pre-trained models are obtained from untrusted sources, assuming the defender has only white-box access to the model and a small, potentially poisoned, unlabeled dataset from the downstream task. The core idea of SCAR is that the "trigger-to-target" backdoor shortcut leaves anomalous signals within the classifier weights for the target class, the model's prediction responses to triggered inputs, and its internal feature representations. Based on this insight, the authors develop two interconnected stages: Class-wise Backdoor Assessment and Shortcut-Guided Model Revision. In the experimental evaluation, the authors test the framework against ten types of backdoor attacks across various model architectures using four datasets—CIFAR-10, CIFAR-100, TinyImageNet, and ImageNet-1k—and compare its performance with several state-of-the-art backdoor detection and mitigation methods.

Critical strengths
  • This paper proposes a lightweight method that does not rely on re-training or trigger inversion.

  • Unified detection and repair framework.

  • Comprehensive empirical performance.

Critical weaknesses
  • Looks like the advantage of SCAR is primarily from the suppression of outputs for target classes, rather than from the accurate identification of backdoor shortcuts.

  • Lack of experiments on clean models.

  • Lack of adaptive attack analysis.

Questions for authors

Thank you for submitting this interesting research to ICLR 2027. I still have a few concerns, which are detailed below:

  1. Computational cost and comparisons with existing detection and defense methods. Although the paper describes the proposed method as a lightweight framework that integrates both backdoor detection and mitigation, providing a more comprehensive comparison of computational costs would further clarify this advantage. For instance, "Channel Lipschitzness-based Pruning" (CLP) [1] mitigates backdoors without requiring additional training. Its channel-wise Lipschitzness statistics could potentially also be used for backdoor detection through an appropriate thresholding strategy.

  2. Evaluation on clean (non-backdoored) models and false positive rates. The paper focuses primarily on detection performance and mitigation efficacy for backdoored models. While clean models are included in the detection evaluation, a more detailed analysis of SCAR’s performance on clean models would be highly valuable. Specifically, how frequently does SCAR misclassify a clean model as a backdoored one? If a clean model is incorrectly flagged, how does the subsequent revision process affect its classification accuracy?

  3. In-depth analytical experiments to explore underlying mechanisms. The paper would be more compelling if it provided further analysis explaining why the proposed "class-wise statistics" and "shortcut-guided revision" are effective. For example, using t-SNE to visualize feature distributions before and after model revision could intuitively demonstrate whether SCAR successfully suppresses responses related to backdoor features while effectively preserving feature representations for benign classes. Furthermore, a more systematic statistical analysis of the "class-specific classifier weights" in both clean and backdoored models would help validate the rationale behind the "static weight anomaly score."

  4. Evaluation against more adaptive and stealthy backdoor attacks. It would be highly valuable to include evaluations against adaptive attacks. For example, Adaptive-Patch[2] and Adaptive-Blend[2] can be thought of as targeting Shortcut Feature Direction Estimation. These types of attacks are designed to make backdoor-related feature representations indistinguishable from benign ones, potentially undermining defense methods that rely on detecting anomalies within the feature space.

[1] Data-free Backdoor Removal based on Channel Lipschitzness, ECCV 2022. [2] Revisiting the Assumption of Latent Separability for Backdoor Defenses, ICLR 2023

5 comments
AuthorsOpenReview ↗

Author Rebuttal Part (1/3)

We thank the reviewer for the constructive feedback and valuable suggestions. We address each of the concerns point by point below.

(1) Shortcut-guided revision vs. target-class suppression.

We agree that it is important to examine whether SCAR’s effectiveness stems primarily from suppressing target-class outputs rather than identifying and mitigating backdoor-related shortcuts. To address this concern, we construct two target-class suppression baselines inspired by prior work on post-hoc logit adjustment [1] and classifier weight scaling [2]: Bias Suppression, which reduces the target-class bias (bc⋆=bc−λb_c^\star = b_c - \lambda), and Weight Scaling, which uniformly scales the corresponding classifier weight (wc⋆=αwc,0<α<1w_c^\star = \alpha w_c, 0 < \alpha < 1). Neither baseline incorporates shortcut-related directions. We evaluate these methods on CIFAR-10 with ResNet-18, adjusting their suppression strengths to achieve comparable clean accuracy before comparing their attack success rates (ASR).

Table 1. Comparison of target-class suppression strategies and SCAR across different backdoor attacks.

AttackBias SuppressionWeight ScalingSCAR
BadNets91.0979.260.13
Blended97.4896.680.01
SSBA99.9199.980.10
WaNet69.2280.182.08

These results demonstrate that direct target-class suppression remains largely ineffective, yielding high ASR across attacks, whereas SCAR substantially reduces ASR at comparable clean accuracy. This advantage stems from SCAR's use of an estimated backdoor shortcut direction to guide targeted logit revision, suppressing trigger-induced predictions while limiting interference with benign predictions. These findings suggest that identifying and mitigating backdoor-related shortcuts, rather than merely suppressing target-class outputs, is key to SCAR's effectiveness. We will include this analysis in Appendix A.1 of the revised paper.

[1] Menon et al., Long-Tail Learning via Logit Adjustment, ICLR 2021.

[2] Kang et al., Decoupling Representation and Classifier for Long-Tailed Recognition, ICLR 2020.


(2) Computational cost and effectiveness comparison.

To assess SCAR's computational efficiency, we compare its execution time with all detection and defense baselines in the main paper, additionally including CLP [3] for defense as suggested by the reviewer. Experiments are conducted under BadNets on CIFAR-10/ResNet-18 (C10) and CIFAR-100/VGG-16 (C100) using a single NVIDIA RTX 3090 GPU. We report execution time averaged over three runs, excluding preprocessing, model loading, and initialization overheads. Tables 2 and 3 present the execution times (in seconds) alongside the corresponding detection and defense performance (%), respectively.

The results show that SCAR achieves the fastest detection on both datasets while maintaining strong detection performance. For backdoor defense, SCAR ranks second only to CLP in execution time. While CLP achieves exceptional computational efficiency through lightweight channel pruning, SCAR achieves lower ASR and higher clean classification accuracy. Overall, SCAR offers a favorable balance between computational efficiency and effectiveness across both tasks. We will include these comparisons of computational efficiency and performance in Sec. 5.3 of the revised manuscript.

Table 2. Comparison of backdoor detection efficiency and effectiveness under BadNets.

MethodC10 Time (s) ↓FPR95 ↓AUROC ↑C100 Time (s) ↓FPR95 ↓AUROC ↑
ABS5.3533.3392.526.1523.8192.97
BAN23.4914.2997.0518.250.00100.00
MM-BD38.0038.1094.78346.780.00100.00
TRODO1.1785.7150.110.5861.9078.00
BARBIE12.2238.1088.4480.5280.9568.48
SPD1.110.00100.004.184.7699.09
SCAR0.080.0099.770.100.00100.00

Table 3. Comparison of backdoor defense efficiency and effectiveness under BadNets.

MethodC10 Time (s) ↓ASR ↓ACC ↑C100 Time (s) ↓ASR ↓ACC ↑
FP80.546.7994.2841.9442.2672.20
NC58.610.6193.7052.600.1669.66
NAD18.231.2892.479.790.0870.18
FT-SAM171.510.9494.3693.040.2366.59
RNP7.571.8990.886.020.0069.71
MM-BM24.900.8694.46149.7220.4272.67
TSBD71.000.8693.7736.840.0469.99
D394.370.1793.8651.560.0171.82
PGBD122.880.7290.8735.950.0161.17
LMR36.630.8494.0391.8611.9970.27
CLP0.630.5792.720.590.7169.14
SCAR2.890.1394.501.680.3273.01

[3] Zheng et al., Data-Free Backdoor Removal Based on Channel Lipschitzness, ECCV 2022.

AuthorsOpenReview ↗

Author Rebuttal Part (2/3)

(3) False alarms on clean models and the impact of revision on accuracy.

In Table 1 of the main paper, we report FPR95, the false positive rate when 95% of backdoored models are correctly detected, providing an initial assessment of SCAR's ability to distinguish clean and backdoored models. Nevertheless, we agree that a separate evaluation of false alarms on clean models and their impact on classification accuracy is valuable. Using the same detection and revision settings as in the main paper, we evaluate SCAR on 50 clean models per dataset, reporting the FPR and the average clean accuracy before and after revision for falsely flagged models only. The results are shown in Table 4.

Table 4. False positive rates on clean models and the impact of subsequent revision on classification accuracy.

Dataset / ModelFalse AlarmsFPR ↓ACC Before ↑ACC After ↑
CIFAR-10 / ResNet-184/508.0094.6794.40
CIFAR-100 / VGG-161/502.0073.6273.54

The results show that SCAR incorrectly flags 4/50 and 1/50 clean models on CIFAR-10 and CIFAR-100, respectively. Importantly, revising these falsely flagged models leads to only marginal average accuracy drops of 0.27 and 0.08 percentage points, respectively. These findings indicate that even when false alarms occur, SCAR preserves benign predictive utility with minimal degradation. We will include this analysis in Appendix A.2 of the revised paper.


(4) Mechanistic analysis of class-wise statistics and shortcut-guided revision.

Following the reviewer’s suggestion, we analyze class-specific classifier weight statistics on CIFAR-10 with ResNet-18 to examine the rationale behind the static weight anomaly score. Specifically, we compare the target-class weight norm with the mean norm of the remaining classes, using the same class as a reference in the clean model. As shown in Table 5, the absolute norm difference is only 0.02 in the clean model, compared with at least 0.19 in backdoored models, approximately 9.5 times larger. The static anomaly score Acst\mathcal A_c^{\mathrm{st}}, defined in Eq. (3) of the main paper, further captures these deviations, assigning substantially higher scores to target classes than to non-target classes across all evaluated attacks. These findings support its effectiveness in highlighting backdoor-related class-wise irregularities.

Table 5. Class-specific classifier weight statistics in clean and backdoored models on CIFAR-10 with ResNet-18.

AttackTarget Class NormNon-target Mean NormNorm Diff.Target Class Acst\mathcal A_c^{\mathrm{st}}Non-target Mean Acst\mathcal A_c^{\mathrm{st}}Score Diff.
No Attack1.901.88+0.020.881.02-0.14
BadNets2.131.83+0.308.141.01+7.13
Blended2.101.82+0.285.860.93+4.93
LC1.611.82-0.215.300.67+4.63
SSBA2.061.80+0.267.961.04+6.92
WaNet2.241.96+0.286.480.98+5.50
LOTUS1.961.77+0.193.180.63+2.55

To illustrate the effect of shortcut-guided revision, we note that SCAR modifies only the classifier head, leaving feature representations unchanged. Thus, feature-level t-SNE visualizations would provide limited insight into its revision mechanism. Instead, we examine the target-class logit margin Mc(x)=fc(x)−max⁡j≠cfj(x)M_c(x)=f_c(x)-\max_{j\neq c}f_j(x) before and after revision across three sample types: triggered, target-class benign, and non-target-class benign. Taking the SSBA attack on CIFAR-10 with ResNet-18 as an example, Table 6 shows that SCAR reduces the average target-class logit margin of triggered samples from 9.07 to −14.16, reversing their preference for the target class and effectively mitigating the backdoor. Meanwhile, benign target-class samples retain positive margins (7.22 to 7.42), and benign non-target-class samples maintain negative margins (−8.93 to −7.27). These results demonstrate that shortcut-guided revision selectively suppresses backdoor-induced predictions while largely preserving benign classification behavior.

Table 6. Average target-class logit margins before and after SCAR revision under SSBA attack on CIFAR-10 with ResNet-18.

Sample TypeBeforeAfter
Triggered9.07-14.16
Target-class benign7.227.42
Non-target-class benign-8.93-7.27
AuthorsOpenReview ↗

Author Rebuttal Part (3/3)

(5) Evaluation on adaptive and stealthy backdoor attacks.

Following the reviewer’s suggestion, we further evaluate SCAR against four adaptive and defense-evasive attacks: Adaptive-Patch and Adaptive-Blend [4], BELT-BadNet [5], and SPD Attack [6]. These attacks are selected to probe the robustness of SCAR’s class-wise assessment and shortcut-guided revision. Specifically, Adaptive-Patch/Blend reduce latent separability through asymmetric trigger planting and correctly labeled trigger-bearing samples; BELT enhances trigger exclusivity to evade defenses relying on approximate triggers; and SPD Attack conceals a deep frequency-domain backdoor behind a shallow decoy backdoor. We evaluate these attacks on CIFAR-10 with ResNet-18, using the same post-training defense baselines as in the main paper. Each entry reports ACC / ASR (%).

Table 7. Comparison of backdoor defense performance under adaptive attacks on CIFAR-10 with ResNet-18.

DefenseAdaptive-PatchAdaptive-BlendBELT-BadNetSPD Attack
No Defense94.06 / 98.8194.29 / 84.2893.56 / 100.0095.21 / 99.99
FP93.80 / 33.9693.90 / 50.5891.51 / 99.7694.12 / 1.31
NC94.06 / 1.6592.96 / 67.9391.10 / 78.4494.03 / 0.74
NAD94.17 / 99.1194.32 / 84.3592.22 / 45.8989.87 / 1.35
FT-SAM93.92 / 75.9694.03 / 41.3792.21 / 82.6394.00 / 0.74
RNP94.06 / 98.8194.29 / 84.2890.06 / 99.2589.98 / 87.54
MM-BM89.06 / 3.3288.50 / 43.6385.26 / 1.5088.14 / 67.00
TSBD94.07 / 98.6293.80 / 62.2191.85 / 10.5793.96 / 1.04
D392.00 / 0.0191.60 / 0.3791.39 / 0.0093.91 / 0.99
PGBD92.51 / 33.6988.95 / 20.5090.71 / 3.2594.21 / 99.99
LMR93.99 / 98.8194.15 / 85.2292.12 / 100.0089.51 / 1.22
SCAR93.56 / 0.0093.29 / 0.0292.58 / 0.0594.57 / 0.00

Table 7 shows that SCAR consistently achieves near-zero ASR across all four attacks, with clean accuracy drops of at most 1.00 percentage point. This robustness can be attributed to SCAR’s complementary class-wise assessment, which captures backdoor-related anomalies at different levels by combining evidence from classifier weights and transformation-induced prediction preferences. Moreover, its shortcut-guided revision targets the underlying trigger-to-target association, enabling effective mitigation without relying on explicit separation of benign and backdoor features or trigger reconstruction. We will incorporate these additional experimental results and analyses into Sec. 5.2 of the revised manuscript.

[4] Qi et al., Revisiting the Assumption of Latent Separability for Backdoor Defenses, ICLR 2023.

[5] Qiu et al., BELT: Old-School Backdoor Attacks can Evade the State-of-the-Art Defense with Backdoor Exclusivity Lifting, IEEE S&P 2024.

[6] Yuan et al., SPD: Shallow Backdoor Protecting Deep Backdoor Against Backdoor Detection, ICCV 2025.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Our Response and Main Updates

Dear Area Chair and Reviewers,

We sincerely thank the Area Chair and Reviewers for their careful evaluation and constructive feedback. In response to reviewers' concerns, we have revised our manuscript with additional experiments, methodological clarifications, and analyses. Below, we summarize the key concerns and corresponding revisions.

High-level summary for the Area Chair.

Our work addresses the challenge of detecting and mitigating backdoors in pretrained classifiers obtained from untrusted sources. We propose shortcut-aware class-wise assessment and revision (SCAR), a unified post-training framework that identifies suspicious target classes through classifier-weight and prediction-response signals, estimates backdoor-related shortcut directions, and performs targeted classifier-head revision without requiring trusted labeled data, trigger reconstruction, or retraining.

Main updates addressing reviewer concerns.

  1. Justification and reliability of shortcut direction estimation (Secs. 4.2, 5.3; Appendix A.1).

    A central concern was whether class-specific weights and feature variability provide a reliable proxy for backdoor-related shortcuts, particularly considering the potential contribution of negative-weight dimensions. We have clarified the proxy's motivation and limitations and provided additional analyses of proxy–oracle shift alignment, oracle-direction alignment, negative-weight contributions, and weight-sign ablations. These findings support the proxy's ability to capture trigger-induced directional information and the empirical effectiveness of prioritizing positive weights, without implying exact shift recovery.

  2. Effectiveness of shortcut-guided revision beyond target-class suppression (Appendix A.1).

    To examine whether SCAR's effectiveness stems merely from suppressing target-class outputs, we have introduced bias suppression and classifier weight scaling baselines, evaluated at comparable clean accuracy. Both alternatives are substantially less effective than SCAR, supporting the importance of shortcut-guided revision beyond direct target-class suppression.

  3. Decision-level effects of shortcut-guided revision (Sec. 5.3).

    Reviewers questioned whether the estimated shortcut direction enables effective revision while preserving benign predictions. To illustrate this effect, we have visualized the target-class margin distributions of triggered, benign target-class, and benign non-target-class samples before and after revision. The results show that revision substantially shifts triggered samples toward negative target-class margins, while benign samples largely retain their original margin characteristics, providing direct evidence of selective backdoor mitigation.

  4. Impact of geometric refinement on shortcut suppression (Sec. 5.3 and Appendix A.1).

    Reviewers questioned whether geometric refinement weakens the shortcut-related direction used for revision. We have examined its alignment with oracle directions and trigger-induced feature shifts under different correction strengths, alongside the corresponding defense performance. The results show that geometric refinement can improve benign accuracy while retaining positive shortcut-related alignment, although excessive correction may weaken backdoor suppression.

  5. Robustness against adaptive and stealthy backdoor attacks (Sec. 5.2).

    To assess robustness against more challenging attacks, we have extended our evaluation to methods that reduce latent separability, enhance trigger exclusivity, or conceal malicious behavior. SCAR maintains effective backdoor mitigation with limited clean accuracy degradation across these settings, providing additional evidence of its robustness against diverse defense-evasive mechanisms.

  6. False alarms and their impact on benign models (Appendix A.2).

    Reviewers raised concerns about the frequency of false alarms and the potential degradation caused by subsequent revision of clean models. We have conducted additional evaluations on clean models and measured the accuracy changes of those incorrectly flagged. The results show relatively low false alarm rates and only marginal accuracy degradation following unnecessary revision.

  7. Computational efficiency and effectiveness (Sec. 5.3 and Appendix A.1).

    To address concerns about computational costs, we have added runtime comparisons with existing detection and defense methods, including CLP as an additional lightweight baseline. SCAR achieves the fastest detection and ranks second in defense efficiency among the evaluated methods, while attaining lower ASR and higher clean accuracy than CLP. These results demonstrate its favorable balance between computational efficiency and defense effectiveness.

We sincerely appreciate the Area Chair's and Reviewers' time and effort in evaluating our work and helping us improve the paper. Thank you for your 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.