acceptodds
Under review as a conference paper at ICLR 2027

Restoring Backdoor Detection under Adaptive Attacks in Multimodal Large Language Models

Abstract

Backdoor detection in multimodal large language models (MLLMs) often relies on predefined statistics of hidden representations or attention. Detector-aware attackers can suppress these statistics during fine-tuning while preserving effective backdoors. We investigate whether detection can be restored after such evasion. In the evaluated settings, defeating an existing detection rule leaves poisoned and benign examples distinguishable under a newly learned score. Motivated by this finding, we propose Benign-Anchored Witness Detection (BAWD), which restores detection by relearning a scoring function on the attacked model. BAWD trains lightweight source classifiers to distinguish audited representations from a trusted same-task benign reference. Given this reference, fitting and thresholding require no additional poisoning labels or exact poisoning rate; thresholding uses a fixed candidate-rank window. Under a matched-reference assumption, the population-optimal source score induces the same ordering as the poisoning posterior, without requiring access to it. Evaluations span three MLLMs and three datasets. On Qwen3-VL-8B fine-tuned on ScienceQA, the linear variant (BAWD-L) achieves 99.8–100.0% detection F1 across seven detector-aware attack configurations, while each of seven baselines degrades substantially under at least one configuration. In two four-round attack sequences targeting previously deployed BAWD detectors, freshly fitted BAWD-L achieves 99.5–99.8% F1 while attack success remains at 99.5–99.8%. These results demonstrate that relearning scores with trusted benign references can restore detection after the evaluated detector-aware attacks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
backdoor defense, poisoned sample detection, adaptive attacks, multimodal large language models, positive-unlabeled learning, data poisoning

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_T95pOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper addresses the issue of MLLM backdoor detection being bypassed by adaptive attacks. The authors propose BAWD - by marking the poisoned training set and the trusted benign reference set separately, training a lightweight classifier to learn the distribution differences between the two, this score is theoretically consistent with the post- poisoning posterior sorting, and then using the maximum score within the fixed ranking window as the cliff threshold. Experiments show that even if the attacker directly targets the deployed detector, the re-fitted BAWD can still maintain an F1 score of over 99.5%, while the baseline detector completely fails.

Critical strengths
  1. The idea of leveraging the fact that the reference set is drawn from the same distribution as the clean data, in order to separate clean from poisoned samples within the audited mixture, is clever and well-motivated.

  2. The paper conducts comprehensive and extensive experiments.

  3. The method is simple in design and has a low deployment barrier.

Critical weaknesses
  1. The paper provides no real, data-driven examples showing that the attacker's processing makes the features hard to distinguish, and that the features become separable again after the defender refits a new witness on the attacked model. Only the manually simulated scatter plots in Figure 1 illustrate this phenomenon.

  2. The filtering ratio is hard-constrained within [3%, 35%], so the method cannot handle cases that fall outside this range. What is the justification for choosing 3% and 35% as the two endpoints?

  3. In practical use, it is difficult to simultaneously obtain the mixed (audited) dataset and a "sufficiently clean and identically distributed" reference set.

  4. The method description in Figure 2 does not correspond one-to-one to the method description in the main text, which increases the reading burden. The six steps shown in the figure should be described separately and aligned with the text.

  5. The paper excessively uses \vspace to compress the page count; for example, there is no vertical spacing between each subsection and its surrounding content.

  6. It is unclear how the proposed method is tied to the multimodal setting. The core mechanism (learning a source-discrimination score between the audited mixture and a clean reference) appears modality-agnostic and would seemingly apply unchanged to unimodal models. The paper would benefit from clarifying what specific challenges the MLLM setting introduces, and what in the method is designed to address them.

Questions for authors

N/A

2 comments
AuthorsOpenReview ↗

Response to Reviewer T95p

We thank Reviewer T95p for the constructive feedback and presentation suggestions. All revisions are highlighted in blue in the updated manuscript, and we respond to each point below.

W1. Real Data-Driven Demonstration of Frozen Failure and Refitted Recovery

To complement the schematic scatter plots in Fig. 1 with real data, we added Fig. 5 in App. M (referenced in Sec. 4.4) using actual cached representations from the Chain-L round-2 checkpoint with MLLM weights held fixed (99.8% ASR). On these identical features, the attacked frozen detector ranks poisoned samples below benign ones (0.2020 AUROC, 0.0% F1), whereas refitting BAWD-L restores clean separation (0.9976 AUROC, 99.5% F1), illustrated both via measured score distributions and shared-coordinate t-SNE projections of all 6,218 audited examples.

W2. Justification and Sensitivity of the [3%, 35%] Candidate Window

The [3%, 35%] topgap endpoints act as a broad engineering range prior within which the sharpest adjacent score drop determines the cutoff: the 3% floor prevents single-outlier cutoffs in the extreme tail, and the 35% ceiling avoids flagging bulk benign data, without fixing the flagged fraction at the true 10% rate. At 10% poisoning, BAWD-L's selected candidate set and F1 are completely invariant under [1%, 20%] and [1%, 50%] across all 11 main settings (App. F.4, Table 16). Below 3% poisoning, the 3% floor mechanically caps precision, but applying a pre-specified [0.5%, 35%] window to the exact same fitted scores without retraining restores high F1 at 2% poisoning (App. K, Table 32):

SettingPoison RateASR (%)AUROCF1 ([3%, 35%])F1 ([0.5%, 35%])
BadNet0.5%89.60.995626.662.9
BadNet1.0%96.50.966834.569.8
BadNet2.0%98.50.998575.597.5
Attention-Spectral2.0%95.60.998873.499.6

Practitioners anticipating lower contamination can therefore simply lower the window floor without retraining.

W3. Practical Availability of a Suitable Reference Set

While cross-task references indeed degrade detection (App. H.4), the reference need only be same-task and mostly clean rather than large, perfectly pure, or drawn from the audited split. When a modest same-task set is retained before adding third-party fine-tuning data, 400 examples already yield 94.4% to 97.9% F1 (Fig. 4a), an ID-disjoint held-out pool yields 99.0% to 99.8% F1 (Sec. 4.6), and 1% to 2% poison contamination still preserves 94.9% to 98.6% F1 on detector-aware attacks while reference contamination stays below the audit rate (Sec. 3.6; Fig. 4b). Furthermore, when no external clean reference exists at all, our new label-free behavioral selection step in App. L (Table 33; Sec. 4.6) builds a same-task pseudo-reference directly from the poisoned audit pool across five draws:

SettingPseudo-Ref AUROCPseudo-Ref F1 (%)Trusted-Ref F1 (%)
Attention-Spectral0.9974 ± 0.000399.70 ± 0.2999.84 ± 0.00
Joint Suppression0.9974 ± 0.000398.85 ± 1.8199.84 ± 0.00
Moment-Gram0.9981 ± 0.000599.66 ± 0.2399.84 ± 0.00
Online-Witness0.9979 ± 0.000699.82 ± 0.0399.84 ± 0.00

W4 & W5. Alignment between Figure 2 and Main Text, and Heading Spacing

We rewrote the Sec. 3 overview to align step-by-step with the six stages of redrawn Fig. 2, progressing from data sources (Step 1, Sec. 3.1), frozen MLLM extraction and shared standardization/PCA (Steps 2 and 3, Sec. 3.3), and source-classifier fitting (Step 4, Sec. 3.4) to OOB aggregation and topgap candidate selection (Steps 5 and 6, Sec. 3.5). We also removed all negative \vspace commands around section and subsection headings to restore the standard ICLR template spacing within the 10-page limit.

W6. Connection to the Multimodal MLLM Setting

While source discrimination itself is a modality-agnostic statistical principle, our audit interface is designed around three challenges specific to generative MLLMs. Because multimodal triggers may reside in the image, the text, or solely their cross-modal conjunction (as in Dual-Key and Adaptive Dual-Key, where BAWD-L achieves 99.5% and 100.0% F1 in Table 1), extraction targets the final prompt token after joint image-text fusion. Stopping before response tokens prevents the detector from directly using repeated target answer text. Class-free source supervision handles free-form MLLM responses without requiring a fixed output taxonomy and also applies to valid-option targets such as B (App. F.3).

We hope these revisions and experiments address all of your concerns, and we would appreciate your considering an updated score.

Reviewer_T95pOpenReview ↗

Response

Thank you for the author's reply. This part has answered my confusion, but I will keep my score.

Reviewer_4sW1OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes BAWD which is a technique to detect backdoors under adaptive attack scenario. They train a classifier on a dataset consisting of clean reference images and an audited set which has a mix of clean and poisoned images, the goal of the classifier is to detect whether the images are from the reference images or the audited images. They train multiple models as this classifier and find it to be highly accurate.

Critical strengths

The paper addresses an interesting and important problem of backdoor detection under the situation where the attacker has a lot of privileged information and they experimental results show very high performance of the classifiers they train.

Critical weaknesses

However the paper has several critical flaws in its current state:

  1. The default poisoning rate is assumed to be at 10%, this is not realistic the practical poisoning rate would be 0.1% or less, the authors need to show the efficacy of their technique at practical poisoning percentages.
  2. In the paper, the experimenters give the defenders a clean reference dataset that is close to the distribution of the audited dataset, but in practical real-world the defender will only be given this audited dataset, how would they construct this clean reference dataset from it (if they could, they already know which datapoints are poisoned, so that's the chicken and egg problem). And if they take any clean reference dataset then what prevents the classifier to latch onto any random feature that differentiates the two datasets (like the classic land vs water bird robustness paper). So the authors need to start with a practical setting where the defender is just given an unclean audited dataset and then show how can they do the backdoor detection.
Questions for authors

Please answer the two points in the weakness section above.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 4sW1

We thank Reviewer 4sW1 for focusing on practical deployment conditions. All new experiments and clarifications are highlighted in blue in the revised manuscript, and we address both concerns below.

W1. Efficacy at Practical Low Poisoning Rates and Scope at 0.1%

While our 10% main-table setting follows recent MLLM backdoor benchmarks such as BYE (NeurIPS 2025, Sec. 5.1, which likewise poisons 10% of ScienceQA with fixed targets) and IAG (CVPR 2026, Sec. 4.1, 5% default), we added low-rate experiments from 0.5% to 2% poisoning in App. K (Table 32), applying both the default [3%, 35%] window and a pre-specified [0.5%, 35%] window to the exact same fitted BAWD-L scores:

SettingPoison RateASR (%)AUROCF1 ([3%, 35%])F1 ([0.5%, 35%])
BadNet0.5%89.60.995626.662.9
BadNet1.0%96.50.966834.569.8
BadNet2.0%98.50.998575.597.5
Attention-Spectral2.0%95.60.998873.499.6

Across these low-rate backdoors, BAWD-L preserves strong poison ranking, and lowering the window floor to 0.5% lifts 2%-poisoning F1 above 97.5% without retraining. We did not evaluate 0.1% because on a 6,218-example instruction-tuning set (unlike web-scale pretraining corpora), 0.1% amounts to roughly six examples, where backdoor implantation is already fragile in our setup (1% Attention-Spectral yields 0.0% ASR, and 0.5% BadNet drops to 89.6% ASR). As App. K states, we make no automated-cutoff claims below 0.5% and recommend using BAWD's ranking to prioritize manual inspection in ultra-low-rate regimes.

W2. Audit-Only Detection without External Clean Data and Preventing Spurious Features

To address the practical "chicken-and-egg" setting where the defender receives only the unverified audited dataset, we removed the external clean-reference assumption in a new end-to-end audit-only pipeline (Sec. 4.6; App. L, Table 33). Starting solely from the poisoned audited dataset and the trained model without any poison labels, we keep each image fixed while reshuffling answer options and substituting questions from other audited rows, scoring how strongly the model still favors the recorded training response. Because fixed-target backdoors continue generating their target regardless of the question whereas ordinary answers change, sampling 1,000 examples from the 30th to 70th percentiles of this behavioral score produces an uncontaminated, same-task pseudo-reference directly from the audit pool (five draws):

SettingPseudo-Ref AUROCPseudo-Ref F1 (%)Trusted-Ref F1 (%)
Attention-Spectral0.9974 ± 0.000399.70 ± 0.2999.84 ± 0.00
Joint Suppression0.9974 ± 0.000398.85 ± 1.8199.84 ± 0.00
Moment-Gram0.9981 ± 0.000599.66 ± 0.2399.84 ± 0.00
Online-Witness0.9979 ± 0.000699.82 ± 0.0399.84 ± 0.00

Mean F1 stays within one point of a ground-truth clean reference across all four checkpoints (input-dependent payloads would require a different selection signal).

Keeping the reference within the same task distribution also directly prevents latching onto spurious features such as image backgrounds in Waterbirds. Spurious shortcuts arise when benign examples differ systematically between the reference and audited sources, which is why cross-task references degrade detection in App. H.4 (Table 29). When the reference matches the task's benign distribution (whether built from the audit pool above or from a small retained same-task split in Sec. 4.6), benign background attributes share identical distributions in both sources, leaving the poisoned subpopulation as the sole systematic discrepancy (Proposition 2). Consistently, applying BAWD-L to a benignly fine-tuned model yields near-chance AUROC (0.53 to 0.59) on trigger-bearing rows (Sec. 4.6), and retraining after removing flagged examples eliminates the backdoor (0.0% ASR) while preserving 97.2% to 98.3% clean accuracy (Table 5), confirming that the classifier isolates the learned backdoor effect rather than benign covariates.

We hope these results address your practical concerns, and we would be grateful if you would reconsider your assessment.

Reviewer_5CZPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the detection of adaptive backdoor attacks against multimodal large language models (MLLMs). The authors observe that while attackers can bypass existing detectors by constraining statistical features such as attention patterns or spectral properties, distinguishable differences often persist between poisoned and clean samples within the model’s hidden representations. To address this, the authors propose the BAWD method, which leverages trusted clean samples as a reference to train a lightweight classifier capable of identifying the anomalous features of poisoned samples. This approach does not require prior knowledge of the trigger, the attack target, or the exact poisoning ratio. Furthermore, the authors introduce an evaluation mechanism called "Witness Chasing" to assess whether a detector can regain its detection capability through retraining after being specifically targeted for evasion by an attacker. Experimental validation on models including Qwen3-VL, LLaVA-NeXT, and InternVL3, across three different datasets, confirms the effectiveness of the proposed method.

Critical strengths

This paper is motivated by the observation that attackers can bypass existing detectors by constraining statistical features such as attention patterns or spectral properties while still maintaining the effectiveness of the backdoor. The authors further discover that such samples often remain separable within the latent space. The design of BAWD is straightforward; it does not require knowledge of the specific trigger or the attack target, needing only a set of trusted, benign samples. Notably, BAWD-L, which employs PCA and a linear classifier, achieves strong results, suggesting that complex detection models are not necessarily required. The experiments encompass a variety of models, datasets, and adaptive attacks, with comparisons drawn against several existing methods. The experiments presented in Table 3, involving sequential attacks against established detectors followed by model refitting—are particularly compelling, as they support the authors' conclusion that detection capabilities can be restored.

Critical weaknesses

BAWD uses trusted benign data to learn a new detector. SCAn and ASSET also use clean reference data, and the source-classification approach is related to PU learning. Although Section 2 discusses these connections, the methodological contribution is still not sufficiently clear to me. Beyond applying the approach to MLLMs and evaluating detection recovery across multiple attack rounds, the authors should clarify what is new in the detector formulation or fitting procedure and why these differences matter. In Table 3, each attack targets the previously fitted detector, which remains frozen during the attack, rather than the process of fitting a new detector. The authors acknowledge this limitation in Section 4.4. They also test an online discriminator and subspace suppression in Appendix G, which provide useful additional evidence. However, these tests do not directly target BAWD’s full refitting procedure. The results show that refitting restores detection under the tested attacks, but it remains unclear how BAWD would perform against an attack specifically designed to evade the newly fitted detector. The method is highly dependent on the quality of the benign reference set. Table 29 shows a significant drop in the F1 score for BAWD-L when cross-task reference data is used, indicating that distribution mismatch severely compromises detection performance. Furthermore, the method inherently requires the filtering of a subset of "suspicious" samples; it cannot directly determine whether an entire dataset is completely clean, which limits its practical applicability.

Questions for authors

Existing works such as SCAn (USENIX Security 2021) and ASSET (USENIX Security 2023) also leverage clean reference data or discrepancies in feature distributions to detect poisoned samples that are difficult to distinguish from clean samples. What are the primary differences between BAWD and these methods? Beyond evaluating detection recovery against adaptive attacks on MLLMs, what other novel contributions does this paper offer in terms of methodological design? The experiments in this paper are primarily conducted at a 10% poisoning rate, while the "topgap" method in BAWD employs a fixed candidate ranking window of 3%–35%. Does the detection performance remain stable when the actual poisoning rate is lower,for instance, at 1% or 2%? The effectiveness of BAWD relies heavily on a trusted, clean reference set; however, in real-world scenarios, it is difficult to guarantee that reference data is entirely clean and follows the same distribution as the data being inspected. As shown in Table 29, the use of cross-task reference data leads to a significant drop in performance. In the absence of trusted, in-distribution clean data, do the authors believe this method still holds practical utility?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 5CZP (Part 1/2)

We thank Reviewer 5CZP for the careful summary and precise questions. Due to OpenReview's per-comment character limit, we post our response in two parts: Part 1 addresses W1, Q1, and W2, and Part 2 addresses Q2, W3, and Q3. All revisions are highlighted in blue in the updated manuscript.

W1 & Q1. Methodological Contribution and Differences from Bagged PU, SCAn, and ASSET

We explicitly credit bagged PU in Sec. 3.4 as BAWD-L's base estimator and do not claim a new PU learning algorithm. Our methodological contribution lies in formulating a class-free, post-attack audit for generative MLLMs, and establishing its recovery under witness chasing (Table 3; App. O), through three design choices: extracting the last prompt token after joint image-text fusion but before response tokens so single- or cross-modal triggers affect the state without exposing target answer text, supervising via reference-versus-audit membership grounded by Proposition 2 without requiring discrete output classes, and selecting candidates via a rate-free topgap rank window (Secs. 3.3 to 3.5).

These choices distinguish BAWD directly from SCAn and ASSET. SCAn estimates within-class variation and applies a cross-class test over discrete output labels, which free-form MLLM targets like Backdoor Attack! lack without either leaking answer text by grouping responses or collapsing the cross-class statistic by pooling. ASSET instead trains an MLP via nested offset optimization to amplify clean-poison variance. Evaluating a matched-feature ASSET adaptation on identical PCA-128 features and 1,000 references across all 11 settings (App. J, Table 31), BAWD-L achieves higher AUROC in every setting and higher F1 than the ASSET adaptation under both reported cutoff rules, without requiring nested adversarial training:

SettingASSET AUROCASSET Code F1ASSET Topgap F1BAWD-L AUROCBAWD-L F1
BadNet0.989428.155.20.999198.5
Blended0.995343.949.90.999099.4
Input-Cond. Freq.0.964954.244.60.991398.1
Dual-Key0.997955.748.90.999999.5
Attention-Spectral0.973643.643.60.998099.8
Joint Suppression0.908542.232.00.997699.8
Moment-Matched0.940557.530.20.998399.8
Moment-Gram0.959643.034.20.997899.8
Online-Witness0.911735.812.40.998499.8
Moment-Matched (B)0.736527.52.20.998099.8
Adaptive Dual-Key0.876142.512.31.0000100.0

W2. Adaptive Attacks Targeting a Newly Fitted Detector

To move beyond attacks on a frozen previous detector, we added a periodically refitted surrogate attack in Sec. 4.4 and App. O (Table 36). Every 50 fine-tuning steps (261 refits per run), the attacker uses its poison labels to refit a BAWD-L surrogate (updating the standardizer, whitened PCA-128 map, and eight class-balanced logistic heads) and backpropagates through its logits with penalty weight WW:

Weight WWASR (%)CP (%)BAWD-L F1 (%)BAWD-L AUROCBAWD-M F1 (%)BAWD-M AUROC
299.696.299.80.998199.00.9993
599.697.999.80.998199.80.9988
1098.997.597.90.998445.60.9912
200.096.87.60.57568.60.6386

Fresh BAWD-L remains accurate across all weights that preserve an active backdoor (W=2,5,10W = 2, 5, 10), whereas W=20W = 20 collapses the backdoor itself. As Sec. 4.4 states explicitly, this surrogate still lacks the defender's private reference and does not differentiate directly through BAWD's refitting procedure, so full bilevel attacks remain open; across all evaluated adaptive strategies, however, we found no configuration that evades fresh BAWD-L while retaining the backdoor.

AuthorsOpenReview ↗

Response to Reviewer 5CZP (Part 2/2)

Q2. Detection Stability at 1% and 2% Poisoning Rates

We evaluated effective low-rate backdoors from 0.5% to 2% poisoning in App. K (Table 32; 1% Attention-Spectral yielded 0.0% ASR and is excluded), applying both the default [3%, 35%] window and a pre-specified [0.5%, 35%] window to the exact same fitted BAWD-L scores without retraining:

SettingPoison RateASR (%)AUROCF1 ([3%, 35%])F1 ([0.5%, 35%])
BadNet0.5%89.60.995626.662.9
BadNet1.0%96.50.966834.569.8
BadNet2.0%98.50.998575.597.5
Attention-Spectral2.0%95.60.998873.499.6

BAWD-L's learned ranking remains stable at low rates (0.9668 to 0.9988 AUROC). Under the default [3%, 35%] window, 2%-poisoning F1 is mechanically bounded near 80% solely because a 3% floor flags at least 187 of 6,218 rows when 124 are poisoned; lowering the pre-specified window floor to 0.5% restores 97.5% and 99.6% F1 at 2% poisoning without retraining.

W3 & Q3. Reference Alignment, Certification Scope, and Utility without External Clean Data

We agree that cross-task references degrade detection because source discrimination captures task shift (Table 29), and that topgap surfaces candidates for inspection rather than certifying a dataset as completely clean (Sec. 3.5). Within this inspection role, removing flagged rows and retraining reduces ASR to 0.0% while preserving 97.2% to 98.3% clean accuracy (Table 5). Crucially, BAWD retains practical utility even without external clean data: in Sec. 4.6 and App. L (Table 33), we construct a same-task pseudo-reference directly from the poisoned audit pool without poison labels by holding each image fixed while reshuffling options and substituting questions (since fixed-target backdoors persist whereas benign answers shift) and sampling 1,000 mid-percentile rows across five draws:

SettingPseudo-Ref AUROCPseudo-Ref F1 (%)Trusted-Ref F1 (%)
Attention-Spectral0.9974 ± 0.000399.70 ± 0.2999.84 ± 0.00
Joint Suppression0.9974 ± 0.000398.85 ± 1.8199.84 ± 0.00
Moment-Gram0.9981 ± 0.000599.66 ± 0.2399.84 ± 0.00
Online-Witness0.9979 ± 0.000699.82 ± 0.0399.84 ± 0.00

Mean F1 stays within one percentage point of using a ground-truth clean reference across all four checkpoints (input-dependent targets would require an alternative selection cue).

We hope these clarifications and new experiments resolve your concerns and support an updated assessment.

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

This paper explores how to defend detector-aware backdoor attacks. Based on the observations, the authors propose Benign-Anchored Witness Detection (BAWD): it uses trusted benign reference samples to train lightweight classifiers which can distinguish poisoned data from benign data through hidden representations of the MLLMs. The method relearns detection scores after adaptive attacks evade existing detection signals. Experiments across three MLLMs and, primarily, ScienceQA, show strong detection capabilities against several adaptive attacks. BAWD can also detect multi-round attacks targeting previously fitted detectors.

Critical strengths
  1. The most important strength, I think, is that the explored problem is meaningful. Currently, backdoor detection methods often rely on statistical metrics or metrics like attention scores, and if attackers have knowledge of such defenses, they can evade these detection methods through special training. It becomes important to know whether we can still defend against backdoors even if the detector is evaded. The authors explore this problem; this is good.

  2. The proposed detection method is easy for implementation. The defender does not need to know the target, trigger, etc. The lightweight classifier is easy for training, and has the potential to be adapted to other scenarios.

Critical weaknesses

Thank you for submitting to ICLR 2027. Here are some critical weaknesses:

  1. The most important weakness I think, is the difference between the proposed method and the traditional Bagged PU learning, also referenced by the authors. Without a direct comparison of BAWD with the prior method proposed in 2014, I am afraid that BAWD looks like an adaptation of PU learning for backdoor detection (lightweight classifiers are very similar). It's OK to get inspired by it; however, the technical novelty of BAWD itself should be further justified. Some special designs, etc.?

  2. Additionally, regarding a previous method, I think it is also important to add specific experiments to compare BAWD with the previous methods. Moreover, the defenders can use reliable reference data, and it's also necessary to discuss some backdoor detection methods that can also use reference data. [1] is an example.

  3. Experiment part. The core results are produced by training on ScienceQA. (leave out examples). This makes me a little worried about data leakage. ScienceQA has some STEM questions that are similar. Maybe you can try another setting, like individually choosing a training and a testing dataset.

A small suggestion is to reorganize the "Problem Setup" Section and make it "Threat Model". This is what a security paper should clearly state. Additionally, redraw the Figures (the current version has so much small text in a large box); you can use AI to assist your drawing, but you may not follow all the styles of it.

You can also add some TSNE visualization to show the classification results.

[1] Pan, Minzhou, et al. "{ASSET}: Robust backdoor data detection across a multiplicity of deep learning paradigms." 32nd USeNIX security symposium (USeNIX security 23). 2023.

Questions for authors

Besides the weaknesses, I also have some questions to ask:

First, I am afraid that the theoretical analysis of this paper is limited. Proposition 1 shows that a separable signal exists; however, BAWD indeed extracts the hidden states, but they may not be the "separable signal" itself. Besides, the gap between theoretical analysis and real problem settings also exists. I think it's OK to clarify the scope of the analysis in the Appendix; however, these things may be weak when regarded as support for the main designs (you list them in the main text).

Second, can the defence method work well when dynamic backdoors towards MLLMs occur? You can refer to [1][2] and other works.

Additionally, could you discuss this situation: the attackers also know the "classifier"-based detection and consider adaptive attacks based on this?

[1] Wang, Zihan, et al. "MTAttack: Multi-Target Backdoor Attacks Against Large Vision-Language Models." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 40. No. 13. 2026.

[2] Li, J., Xu, B., Chen, S., Li, J., Lei, J., Zhao, H., & Zhang, D. (2026). Iag: Input-aware backdoor attack on vlm-based visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 27872-27883).

2 comments
AuthorsOpenReview ↗

Response to Reviewer NZww (Part 1/2)

We sincerely thank Reviewer NZww for the detailed review and concrete suggestions. Due to OpenReview's per-comment character limit, we post our response in two parts: Part 1 addresses W1 to W5, and Part 2 addresses Q1 to Q3. All revisions are highlighted in blue in the updated manuscript.

W1. Connection to Bagged PU Learning and Methodological Novelty

We agree that BAWD-L's base classifier follows Mordelet & Vert (2014). Sec. 3.4 now credits bagged PU explicitly, retaining unlabeled subsampling and out-of-bag aggregation with class-balanced logistic regression rather than claiming a new PU estimator. Standard PU assumes static distributions, whereas a detector-aware attacker reshapes hidden states to suppress the deployed score. Our contribution lies in studying post-attack recovery via witness chasing (Secs. 3.2 and 4.4) and tailoring an MLLM audit around this estimator: extracting pre-response joint prompt states so cross-modal triggers affect the score without leaking target answer text, supervising by class-free source membership, and thresholding via a rate-free topgap window (Secs. 3.3 to 3.5; ablations for B=8B = 8 vs. 11 and PCA in Table 4).

W2. Comparison with Reference-Based Defenses (ASSET)

Following your suggestion, we added a matched-feature adaptation of ASSET in App. J (Table 31) on the exact same cached PCA-128 features and 1,000 benign references as BAWD-L, retaining ASSET's nested selection, variance loss, and released threshold (Code F1), alongside topgap on the same ASSET scores:

SettingASSET AUROCASSET Code F1ASSET Topgap F1BAWD-L AUROCBAWD-L F1
BadNet0.989428.155.20.999198.5
Blended0.995343.949.90.999099.4
Input-Cond. Freq.0.964954.244.60.991398.1
Dual-Key0.997955.748.90.999999.5
Attention-Spectral0.973643.643.60.998099.8
Joint Suppression0.908542.232.00.997699.8
Moment-Matched0.940557.530.20.998399.8
Moment-Gram0.959643.034.20.997899.8
Online-Witness0.911735.812.40.998499.8
Moment-Matched (B)0.736527.52.20.998099.8
Adaptive Dual-Key0.876142.512.31.0000100.0

BAWD-L achieves higher AUROC in all 11 settings and higher F1 than the matched-feature ASSET adaptation under both its released threshold and the common topgap rule. Unlike class-conditional methods such as SCAn (Sec. 2), which require discrete output classes absent in free-form MLLM targets, ASSET provides a direct baseline on shared features.

W3. Question Similarity in ScienceQA and Separate Datasets

Similar STEM questions do not cause train-test leakage because BAWD audits the fine-tuning set transductively via out-of-bag scoring without poison labels (Sec. 3.5), rather than training on one split and testing on another. Since the benign reference holds zero poisons, question similarity reveals nothing about trigger presence, as confirmed by near-chance AUROC (0.53 to 0.59) on a clean model (Sec. 4.6) and 99.0% to 99.8% F1 with an ID-disjoint reference (App. D.1). Moreover, Table 2 independently evaluates the full pipeline on PhD and SEED-Bench, reaching 98.9% to 100.0% F1.

W4 & W5. Threat Model, Redrawn Figures, and t-SNE Visualization

We reorganized Sec. 3.1 into an explicit Threat Model, redrew Figs. 1 and 2 with concise labels, and added Fig. 5 (App. M) showing ground-truth labels and BAWD-L outcomes for all 6,218 audited examples from Chain-L round 2 on shared t-SNE coordinates, marking the four false positives and two missed poisons alongside measured score distributions.

AuthorsOpenReview ↗

Response to Reviewer NZww (Part 2/2)

Q1. Scope of Theoretical Analysis (Proposition 1)

We agree that Proposition 1 is an existence result rather than a guarantee for last-prompt-token states, and Sec. 3.2 (App. B) frames it strictly as motivation: a nonzero ASR-BTR gap implies trigger-related separation in any representation sufficient for the target output, so defeating one fixed statistic need not erase internal evidence. Whether our extracted states retain this separation is verified empirically in App. G.2 (Fig. 3b), where fresh linear probes achieve AUROC of at least 0.990 at every layer from layer 20 to the final layer across all rounds of both chains.

Q2. Dynamic and Multi-Target Backdoors (MTAttack and IAG)

Source supervision assumes neither a static trigger nor a single target. For input-aware dynamic perturbations like IAG, Table 1 already evaluates an input-conditioned U-Net frequency trigger bounded by 8/255, where BAWD-L reaches 98.1% F1 (reproducing IAG directly requires a bounding-box grounding pipeline outside our text-generation setup). To test dynamic multi-target attacks directly, we added a three-target MTAttack adaptation on LLaVA-NeXT-8B and ScienceQA using the official trigger-optimization code (Sec. 4.2; App. N, Tables 34 and 35; clean accuracy 92.8%):

TargetAudited PoisonsASR (%)Cross-Target Conf. (%)BAWD-L Recall (%)BAWD-M Recall (%)
Backdoor Attack!20886.99.498.699.0
Access Denied.20794.35.398.699.0
System Error.20788.49.899.5100.0
Method (Pooled Audit)Precision (%)Recall (%)F1 (%)AUROCAUPRCFlagged
BAWD-L99.798.999.31.00000.9998617
BAWD-M93.299.496.20.99920.9935663

All three targets retain high ASR while the pooled audit recovers poisons uniformly across all three trigger-response pairs.

Q3. Classifier-Aware Adaptive Attacks

Table 3 already grants the attacker white-box access to the deployed BAWD classifier (standardizer, PCA, and all heads), yet refitting BAWD-L restores 99.5% to 99.8% F1 at 99.5% to 99.8% ASR. Going further, App. O (Table 36) tests an attacker that supervisedly refits a BAWD-L surrogate (standardizer, PCA-128, and eight logistic heads) every 50 steps (261 refits per run) and backpropagates through its logits with penalty weight WW:

Weight WWASR (%)CP (%)BAWD-L F1 (%)BAWD-L AUROCBAWD-M F1 (%)BAWD-M AUROC
299.696.299.80.998199.00.9993
599.697.999.80.998199.80.9988
1098.997.597.90.998445.60.9912
200.096.87.60.57568.60.6386

Fresh BAWD-L remains effective across all runs that preserve an active backdoor (W=2,5,10W = 2, 5, 10), whereas W=20W = 20 eliminates the backdoor itself. Sec. 4.4 explicitly notes that bilevel optimization directly through private-reference refitting remains open.

We hope these new experiments and revisions address all of your concerns, and we would appreciate your considering an updated assessment.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We sincerely thank the Area Chair and all reviewers for their thoughtful feedback and constructive suggestions. All revisions in the updated manuscript are highlighted in blue. Below, we summarize our key contributions in the paper, followed by the additions and clarifications made in response to the reviews.

Existing poisoned-sample detectors for multimodal large language models (MLLMs) largely rely on predefined statistics of hidden representations or attention, which detector-aware attackers can suppress during fine-tuning while preserving effective backdoors. Our work shows that suppressing a monitored statistic, or even a previously deployed detector, leaves poisoned and benign examples distinguishable in the attacked model's hidden states under a newly learned score. Building on this finding, we propose Benign-Anchored Witness Detection (BAWD), a post-attack audit built around a bagged-PU source classifier that contrasts pre-response joint prompt representations against a same-task benign reference and selects candidates via a rate-free rank window, without requiring output-class partitions or updates to the audited model. Across three MLLMs, three datasets, and two four-round witness-chasing chains targeting deployed detectors, refitting BAWD-L consistently restores poisoned-sample detection in settings where the evaluated predefined-statistic detectors fail.

To address specific requests and provide further clarity, we have included the following additions and updates in the manuscript:

  • Section 1 and Section 3.4: Explicitly credit bagged PU learning as BAWD-L's underlying estimator, clarify BAWD's MLLM-specific audit design, and scope our third contribution to the evaluated predefined-statistic detectors.
  • Section 3 and Figures 1 and 2: Reorganize Section 3.1 as a formal Threat Model, align the Section 3 overview step-by-step with redrawn Figures 1 and 2, and restore standard heading spacing throughout the paper.
  • Appendix D.1: Clarify that the held-out disjoint reference control separates sample identities rather than ruling out content-level similarity.
  • Appendix J (Table 31): Add a matched-feature adaptation of the reference-based defense ASSET on identical PCA-128 representations and 1,000 benign references across all 11 main settings.
  • Appendix K (Table 32) and Section 4.5: Evaluate low poisoning rates from 0.5% to 2% under the default [3%, 35%] window and a pre-specified [0.5%, 35%] sensitivity window, and state the scope boundary below 0.5%.
  • Appendix L (Table 33) and Section 4.6: Introduce a label-free behavioral selection procedure that constructs a same-task pseudo-reference directly from the poisoned audit pool without external clean data.
  • Appendix M (Figure 5) and Section 4.4: Add measured score distributions and shared-coordinate t-SNE outcomes on fixed MLLM weights from Chain-L round 2 to illustrate evasion and recovery on real data.
  • Appendix N (Tables 34 and 35) and Section 4.2: Evaluate a three-target adaptation of the dynamic multi-target attack MTAttack on LLaVA-NeXT-8B and ScienceQA.
  • Appendix O (Table 36) and Section 4.4: Evaluate attacks with a supervised surrogate refitted every 50 training steps, and state explicitly that bilevel attacks optimizing directly through BAWD's private-reference refitting 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.