- 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
-
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.
-
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
-
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.
-
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
-
What is the specific experimental setup for Figure 1, and how are the quantities on the horizontal and vertical axes defined?
-
How is the filtering threshold in line 306 selected, and how sensitive are the results to it?
-
Why does SCAR have relatively high FPR95 values for LC and SSBA on CIFAR-10 in Table 1?
-
Do the baselines also use unlabeled defense data containing poisoned samples? How are baselines that require labels handled?
8 comments
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 as the centered directional response and as the resulting logit correction. Since only the target-class logit is modified, the revised margin satisfies , where . For triggered inputs predicted as class , successful suppression requires to reverse their target-class preference. Conversely, preserving benign predictions of class requires . For benign inputs predicted as other classes, a positive correction further reduces the target-class logit without weakening their decision margins against class . 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 guarantees only a zero-mean correction over the reference samples predicted as class , 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 ( and ). Moreover, Table 6 of the main paper shows that the combined direction achieves an ASR of , compared with using the shortcut-related direction alone, while improving ACC from to . 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 to 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.
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 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 , 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 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 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 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.
| Quantile | ACC ↑ | ASR ↓ |
|---|---|---|
| 0.90 | 94.64 | 4.00 |
| 0.91 | 94.27 | 1.19 |
| 0.93 | 94.50 | 0.13 |
| 0.94 | 94.35 | 1.02 |
| 0.96 | 93.71 | 3.17 |
(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.
| Method | Ground-truth labels | Trusted clean data |
|---|---|---|
| FP | ✓ | ✓ |
| NC | ✓ | ✓ |
| NAD | ✓ | ✓ |
| FT-SAM | ✓ | ✓ |
| RNP | ✓ | ✓ |
| MM-BM | ✓ | ✓ |
| TSBD | ✓ | ✓ |
| D3 | ✓ | ✓ |
| PGBD | ✓ | ✓ |
| LMR | ✓ | ✓ |
| SCAR | ✗ | ✗ |
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 sc and its oracle counterpart scora 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.
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.