acceptodds
Under review as a conference paper at ICLR 2027

Backdoor Success Outside the Data Manifold: A Geometric Study of CLIP Backdoor

Abstract

Contrastive Language–Image Pretraining (CLIP) has substantially advanced vision–language representation learning through self-supervised pretraining on web-scale image–text pairs. However, its heavy reliance on uncurated web data leaves it highly susceptible to backdoor attacks. Existing CLIP backdoor attacks can achieve high attack success rates, but they often leave a detectable geometric trace, where backdoor representations are pushed outside the clean data manifold, making them vulnerable to representation-based defenses. We mainly attribute this to a pervasive geometric phenomenon across CLIP backdoors, named backdoor representation off-manifoldness. In this work, we trace the phenomenon to the modality gap between CLIP's image and text embedding spaces because cross-modal backdoor objective pulls a triggered image(text) representation toward a textual(image) target, and a large fraction of the backdoor gradient lies outside the subspace supported by image(text) representations. Guided by this insight, we propose the On-Manifold Backdoor Attack (OMA), which retains the standard cross-modal backdoor objective while regularizing triggered representations against an intra-modal target feature bank via a principal-component normal loss and a local mean-squared-error loss. Across ImageNet-1K, ImageNet-R, and ImageNet-Sketch, OMA maintains a near- undefended attack success rate while effectively suppressing anomaly detection signals. Our study connects CLIP backdoor detectability directly to the modality gap and demonstrates that attack success can be decoupled from off-manifold detection signatures. Our code is in the supplementary material and will be open-sourced upon acceptance.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Backdoor attack, Self-supervised learning, CLIP

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_5xDMOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper focuses on the creation of backdoor or poisoned examples. It makes the observation that classic ways of creating backdoor samples result in ones that lie off the data manifold and are hence potentially vulnerable to detect via geometric measures. Leveraging this observation they proposed a new method to create on maniold backdoor examples that are harder to detect.

Critical strengths

-makes an intriguing observation that conventional backdoor samples are geometrically vulnerable to detection, opening the door to manifold manipulation techniques to improve stealthiness

-new loss function proposed to keep backdoor samples on manifold follows logically from collection of evidence and step by step observations

-experimental results are strong

Critical weaknesses

-proposed new loss function requires two hyperparameters – it is not clear how to choose these and how sensitive resulting behaviour is – please elaborate

Questions for authors

-the geometric measures (k-nn distance, DAO score) are estimates of population quantities. Are any of the conclusions in the paper affected by estimation error for these geometric measures? -does the use of pca in formulating the revised loss function eq 5 mean there is an implied “linearity” assumption?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 5xDM (2/2)

(Part 2/2)

The results show three points:

  • The absolute values of the detection metrics depend on NN, but the comparison between clean and backdoor samples does not. Raw scores carry a bias shared by clean and backdoor samples. For example, when NN increases from 128 to 32,768, the mean kkNN distance of BadCLIP's clean queries drops from 0.904 to 0.642 and that of its backdoor queries drops from 1.049 to 0.924, but the backdoor queries are farther at every NN.
  • The conclusion that cross-modal attacks are off-manifold is robust, and a finite reference bank only underestimates it. For CLIP-Backdoor and BadCLIP, all metrics already give AUROC ≥ 0.92 at N=128N=128 and ≥ 0.97 for N≥1,024N\ge 1{,}024, with std ≤ 0.013 across random banks. AUROC increases with NN: the more accurate the estimate, the more clearly backdoor samples are separated. The same holds for the target-class local PCA residual (AUROC between 0.990 and 0.994 as the target bank grows from 128 to 1,000).
  • The conclusion that OMA reduces off-manifoldness likewise does not depend on the reference bank; only the DAO metric is an exception. For OMA, the AUROC of the kkNN distance, PCA residual, and normal residual stays below 0.42 at every NN, and the target-class local PCA residual stays at 0.29–0.33. DAO is the only metric sensitive to NN: OMA's DAO AUROC rises from 0.25 to 0.58–0.72 for N≥1,024N\ge 1{,}024, with std up to 0.15 across reference banks. The in-modal MSE control shows the same trend (0.03 → 0.53), indicating that this is because the DAO estimator (local intrinsic dimension and kk-th-nearest-neighbor radius ratio) has high variance on on-manifold samples, rather than a property specific to OMA. Even at its maximum, it remains far below the ≥ 0.99 that DAO gives for cross-modal attacks. The DetectCLIP AUROC of OMA reported in the paper (59.51% on ImageNet-1K) also falls within this range.

We also bootstrapped the query samples at N=2,048N=2{,}048 to measure the query-side sampling error: the half-width of the 95% confidence interval of AUROC is at most 0.01 for cross-modal attacks and at most 0.06 for OMA and MSE, both much smaller than the gap between the two groups.

3. Whether using PCA in Eq. (5) implies a "linear" assumption (Question 2)

You raised the concern:

Does using PCA in the revised loss function in Eq. (5) imply some kind of "linear" assumption?

No. We implement the normal loss with PCA decomposition because it is simpler to implement. PCA is a commonly used technique in machine learning: it has a closed-form solution and requires few design choices, which benefits reproducibility and interpretability. This choice is common in the deep learning literature [1, 2], and it is not meant to assume that the representation space has any linear structure.

In fact, in Section 5 of the original paper (Anchoring to populated target regions), we also explained why we did not use non-linear distribution matching:

Ideally, OMA would use a distribution-matching loss that makes the marginal distribution of triggered representations indistinguishable from that of clean target representations, as illustrated in Figure 3(e). This is the ideal loss because it directly encodes the desired on-manifold outcome, rather than merely constraining individual representations to a subspace or prototype. In practice, estimating and optimizing such a loss from a small poisoned set is difficult. We therefore use a pointwise surrogate that is stable under limited data.

An exact non-linear distribution-matching loss would have to estimate a high-dimensional distribution from only 500 poisoned pairs (a 0.1% poisoning rate), which is highly unreliable, whereas the PCA-based normal loss works well with the few poisoned samples available. The PCA subspace is only a coarse approximation of the clean support. It retains 95% of the target-class variance, so even if the clean manifold is curved, the subspace approximately contains it. Penalizing the component outside this subspace is therefore a necessary condition for being on-manifold, not an assumption that the manifold is flat.

We hope the above clarifications address your questions.

[1] Haoyu Ma, Xiangru Lin, Yizhou Yu. I2F: A Unified Image-to-Feature Approach for Domain Adaptive Semantic Segmentation. TPAMI, 2024.

[2] Luo et al. Robust Graph Condensation via Classification Complexity Mitigation. NeurIPS, 2025.

Area_Chair_VJETOpenReview ↗

Please engage in the discussion with authors

Dear Reviewer 5xDM,

Please engage in the discussion with authors, thank you!

AC

Thankyou, acknowledging your response.

Reviewer_JKv8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the geometric properties of backdoor representations in CLIP and argues that conventional cross-modal backdoor attacks often produce off-manifold representations due to the modality gap. Based on this observation, the authors propose OMA, which encourages backdoor representations to remain close to the clean target-class distribution.

Critical strengths

The paper investigates an interesting problem from a geometric perspective and provides an intuitive explanation for the detectability of existing CLIP backdoor attacks. The analysis of representation geometry is also potentially useful for understanding backdoor behavior in multimodal models.

Critical weaknesses

My main concern is that the proposed attack may primarily exploit the limitations of existing anomaly detectors rather than actually making backdoor representations indistinguishable from benign ones. In particular, some reported AUROC values are substantially below 50%, even approaching zero. Such a low AUROC does not necessarily indicate that benign and backdoor representations are difficult to distinguish. Instead, it may suggest that the anomaly scores are highly discriminative in the opposite direction. One possible explanation is that the proposed geometric regularization pushes backdoor representations excessively close to clean target samples, creating representations that are unusually "normal" according to certain one-sided anomaly scores. These representations may evade detectors that only identify large deviations while remaining statistically distinguishable from genuine benign samples. Hence, this paper should provide more adaptive evaluations to support their core claim, i.e., crafting "normal" backdoor representations to evade detectors even under adaptive defenders who knows this attack strategy.

Questions for authors

Could the authors evaluate the proposed method against adaptive defenders, such as detectors with recalibrated anomaly scores (evaluate on more robust adaptive detectors if possible) , to demonstrate that the backdoor representations are genuinely indistinguishable from benign samples rather than merely exploiting the one-sided detection criteria of existing methods?

Reviewer_b95iOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper observes that several CLIP backdoor attacks push triggered features away from the clean data distribution and links this behavior to the gap between image and text features. Based on this analysis, the authors propose OMA. It keeps the original cross-modal backdoor objective and adds a PCA-based normal-direction constraint and a local MSE loss, using clean target-class features as a reference. These losses reduce feature shifts in abnormal directions and bring backdoor features closer to clean target samples. Experiments cover image-side and text-side attacks on ImageNet-1K, ImageNet-R, and ImageNet-Sketch, measuring attack success, clean performance, and resistance to several defenses. OMA maintains high attack success without defenses and reduces the effectiveness of some geometric detectors.

Critical strengths

1.The paper asks a clear question about the link between backdoor feature geometry and detectability. The connection between the modality gap and the attack design is easy to follow. 2.OMA follows naturally from the geometric analysis. The ablations also examine the roles of its two added losses. 3.The experiments cover several attacks, detectors, and repair methods, while reporting both attack success and clean performance.

Critical weaknesses

1.The main results use an older CLIP model, leaving the security impact on current multimodal systems unclear. The attack experiments mainly use the original CLIP ViT-B/16 trained on a CC3M subset, with classification and retrieval evaluations. The SigLIP analysis in Figure 2 and the analyses of other encoders in the appendix do not test OMA attacks on those models. The geometric study has value, but it remains unclear whether the attack survives newer training objectives, encoders, or downstream adaptation. Experiments with a newer encoder, such as SigLIP 2, continued training, or integration into a multimodal large language model would better support its practical relevance.

2.OMA's practical advantage against existing defenses remains limited. Appendix C, Table 5 shows that Decomp detects image-side OMA on all three datasets, with both F1 and TPR@5% reaching 100%. The authors should explain why Decomp still detects these attacks after OMA reduces their geometric anomalies. They should also clarify what additional risk OMA poses compared with earlier attacks when Decomp, or a combination of existing detectors, is used. These results should be discussed in the main text.

3.The main result tables are difficult to read. Tables 2 and 3 contain too many metrics and values in a very small font. The authors should improve their readability and check the font size against the conference formatting requirements.

Questions for authors

The research question is reasonable, and the geometric explanation connects well with the OMA design. I therefore recognize the value of the mechanism study. However, the main experiments remain limited to classification and retrieval with an older CLIP model. The impact on current multimodal systems is unclear, and Decomp already detects the tested image-side OMA attacks. I would consider raising the score if the rebuttal provides convincing evidence or clarification about newer models and downstream applications, and explains the attack's limits against existing defenses.

Comments on the paper

1 comment
Area_Chair_VJETOpenReview ↗

Follow-Up on the Authors’ Rebuttal

Dear reviewers,

The authors have posted their rebuttal. Please take a moment to read their response and leave a brief follow-up comment on whether it addresses the concerns you raised in your review. If any points remain unclear or you have follow-up questions, please raise them during the rebuttal period so the authors have a chance to respond. Please also indicate whether the rebuttal changes your overall assessment of the paper.

Thank you !

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.