acceptodds
Withdrawn from ICLR 2027

From Rigid Carriers to Flexible Wrappers: Towards Output-Concealment Shadow-Activated Backdoor Attacks on MLLMs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

BadMLLM introduces the ***S***hadow-activated ***B***ackdoor ***A***ttack (***SBA***) for MLLMs. Malicious responses are triggered when generated content mentions a predefined target entity, rather than by explicit input triggers. (1) However, existing attacks, including SBA, neglect output-side concealment. Their rigid response patterns remain detectable by output-analysis defenses. This paper decomposes the backdoor carrier, i.e., the full sentence containing the malicious goal, into an invariant goal and flexible wrappers. The decomposition shifts similarity of backdoor carriers from sentence-level to lexical-level. (2) Unfortunately, directly applying Supervised ***F***ine-***T***uning (***SFT***)-based methods, such as BadMLLM, fails to lead victim MLLMs to associate entity mentions with goal insertion under varied wrappers. To resolve this tension, we formulate backdoor implantation as a constrained multi-objective optimization problem. Leveraging the flexible reward-driven optimization of ***R***einforcement ***L***earning (***RL***), a diverse-Wrapper SBA framework (***WrapperSBA***) is proposed to jointly optimize attack effectiveness, wrapper diversity, positional variation, and response quality. A sample-level instruction decay mechanism further addresses the association difficulty and gradually aligns training with prompt-free inference. (3) Conventional RL reward designs use fixed weights as hyperparameters to balance different reward components, which can cause reward hacking by overfitting to individual components. This paper introduces a dynamic weight scheduler that adaptively recalibrates coefficients according to measured contributions. Experiments show that WrapperSBA increases ASR from 46.57% to 99.40% and substantially reduces Self-BLEU of wrappers from 1.00 to 0.40. WrapperSBA substantially reduces the detection effectiveness of the evaluated defenses relative to fixed-wrapper SBA by enhancing output-side concealment. Further validation across unimodal settings and diverse attacks supports the defense-evasion benefit of wrapper diversity.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Backdoor Attack, Backdoor Defense

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_UzsWOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the output side of shadow-activated backdoor attacks (SBA) on MLLMs, in which malicious content is generated whenever the model’s own response mentions a target entity. The authors decompose the backdoor carrier into an invariant attack goal (a fixed malicious URL) and a flexible linguistic “wrapper.” They argue that SFT cannot learn this decomposition, because wrapper diversity removes the shared surface form that makes the entity→goal association learnable. To address this, they formulate backdoor implantation as a constrained multi-objective GRPO problem (WrapperSBA), with rewards for attack efficacy, semantic and positional wrapper diversity, relevance, and fluency.

Critical strengths

The goal–wrapper decomposition (Sec. 1, Fig. 1) is a clean reframing of output-side stealth, and it is actually operationalized rather than left as a slogan: the invariant goal is enforced by r_core (Eq. 4), while diversity is rewarded on the extracted carrier sentence (Eq. 5). The SFT-vs-RL comparison in Table 1 is the right ablation for the paper’s central claim. The w/o-RL variant achieves comparable Self-BLEU (0.51–0.57) but its ASR drops to 46.57–85.49, which isolates the claimed tension between diversity and activation. Appendix B describes each defense re-implementation concretely (PurMM’s forward hook on mm_projector, ToT’s TSC+CDG conjunction with a τ sweep, BYE’s bimodality index cutoff, SRD’s SFS with α=0.5). This is more operational detail than attack papers usually provide for the defenses they claim to break. Table 2’s sweep of goal length L against ConfGuard’s window and the sampled generations in Appendix C (Tables 11–13) are honest. They show exactly where evasion fails and let readers judge the diversity claim directly, rather than relying on Self-BLEU alone.

Critical weaknesses

The central claim that WrapperSBA “substantially reduces the detection effectiveness of the evaluated defenses” is contradicted by the paper’s own data. In Table 4, the RL variant has an average ToT TPR of 66.87 versus 19.60 for BadMLLM, i.e. it is detected far more often; CleanGen retains 91–100% TPR under diverse wrappers (Table 9); and SRD/BYE are at chance level for every method (SRD: TPR 9.13 vs FPR 11.73 for RL, 9.87 vs 10.67 for BadMLLM), so these comparisons cannot support any causal claim. The claims should be narrowed to confidence-pattern defenses such as ConfGuard, and detectors should be compared by ROC/AUC rather than at a single author-chosen operating point. The ConfGuard evasion is attributed to wrapper diversity, but Table 2 shows that it is governed by payload length relative to the sliding window: TPR rises to 94.1/100 once L>window. Moreover, the unimodal URL was explicitly chosen as 9 tokens against a window of 10 (Appendix B.2). The active variable is therefore payload length, not diversity. A clean experiment would hold L fixed while varying only wrapper diversity, and would report ConfGuard with a defender-adapted window size. Headline numbers are assembled from different rows, and the key ablation comparison is unfair. The abstract’s “46.57%→99.40%” and “Self-BLEU 1.00→0.40” mix datasets and model sizes; Sec. 4’s “Self-BLEU 0.40 and position entropy 0.81” come from Laptop-13B and Sandwich-13B respectively; and the 38.72-point gain in Sec. 4 compares the selected checkpoint against epoch 14 of the no-scheduler run, an epoch that Alg. 1’s a_t+n_t checkpoint rule would never select. Please report per-setting means ± std over multiple seeds and compare best-selected checkpoints on both arms. No numeric hyperparameters are given anywhere for the reward or the scheduler (w_succ, w_fail, w_noent, w_sil, w_url, w_glb, τ_dup, τ_rel, τ_sat, [s_min,s_max], γ, ζ, ν_c, λ_p, R_max, group size K, learning rate, LoRA configuration for the MLLM runs, compute budget), and Appendix A.3 restates Alg. 2 in prose instead of supplying them. This makes the Reproducibility Statement’s claim that the details are in the text inaccurate. A single table of all reward and scheduler constants would fix this within the page limit.

Questions for authors

See weaknesses

3 comments
AuthorsOpenReview ↗

Response to Weakness 1-3

We would like to sincerely thank the reviewer for their valuable comments and suggestions, which help us significantly improve the quality of our paper. All modifications in the new version of the paper are marked in blue. We sincerely hope that our response can address your concerns.

Weakness 1. Scope of Defense Claims

A defense is considered effective only when it achieves both a high true positive rate (TPR) and a low false positive rate (FPR).Our intended interpretation is that, compared with the fixed-wrapper BadMLLM variant, WrapperSBA generally induces higher TPRs but also substantially higher FPRs. This distinction is important because the increased TPR does not necessarily indicate improved detectability in practice, as it comes at the cost of misclassifying more benign samples as malicious.Specifically, BadMLLM exhibits relatively low TPR and low FPR, whereas WrapperSBA tends to induce higher TPR and higher FPR. This suggests that WrapperSBA makes existing detection methods more likely to flag benign responses, thereby preventing them from simultaneously achieving effective attack detection and an acceptable false alarm rate. In scenarios where detection fails, comparing which method evades detection more effectively based on a single metric alone is uninformative. Table 4 shows that, for WrapperSBA (with RL), ToT detects more triggered responses, with an average TPR of 66.87% compared to 19.60% for BadMLLM. However, its average FPR also increases from 29.53% to 60.73%. Therefore, these results indicate a different detection profile rather than a degradation in detection capability.

Weakness 2 : Operating points and ROC/AUC evaluation

We agree that TPR alone is insufficient. Our tables report FPR together with TPR, and the ToT, Probe, BYE, and SRD tables additionally report F1 and F0.5. ROC/AUC would provide complementary threshold independent information, but it is not essential. TPR, FPR, F1, and F0.5 are sufficient for our evaluation. Additional redundant metrics would unnecessarily bloat the tables. This reporting practice has direct precedents in the evaluated literature. ConfGuard, AAAI 2026 (https://arxiv.org/html/2508.01365v1) evaluates TPR and FPR at explicit probability and length thresholds and provides separate threshold ablations rather than relying on ROC/AUC. STRIP, ACSAC 2019 (https://arxiv.org/html/1902.06531) reports the false acceptance rate at a preset false rejection rate of 1 percent. CleanGen, EMNLP 2024 (https://aclanthology.org/2024.emnlp-main.514/) likewise evaluates defense behavior using task and detection metrics without making ROC/AUC a prerequisite. These examples do not diminish the value of ROC/AUC, but they show that paired TPR and FPR at documented operating points are established evaluation practices.

Weakness 3:Question about ConfGuard

ConfGuard validates a specific mechanism of wrapper diversification. ConfGuard's detection performance depends strongly on both the target URL length and its sliding-window configuration. We optimize the paper by designing the new Figure 3 to replace the Table 2, reporting "Diverse wrappers benefit the stealthiness of WrapperSBA by shrinking the fixed carrier below ConfGuard’s window". Specifically, let LcL_c denote the length of the complete fixed carrier and Lg<LcL_g<L_c the length of its invariant payload. BadMLLM exposes a shared high-confidence region of length LcL_c, whereas WrapperSBA reduces the shared region to LgL_g and diversifies the remaining wrapper. WrapperSBA obtains 4.8–22.8% TPR for short cores , compared with 100% for fixed-wrapper SBA. Reducing the detection window introduces substantial practical trade-offs. Natural responses frequently contain expressions similar to the invariant payload, making smaller windows increasingly susceptible to false positives. Furthermore, shorter windows substantially increase computational overhead due to the larger number of candidate segments that must be examined. Therefore, although ConfGuard can potentially detect shorter invariant carriers by reducing its window size, doing so comes at the cost of increased false-positive rates and computational complexity. This further demonstrates the practical value of wrapper diversification.

AuthorsOpenReview ↗

Response to Weakness 4

Weakness 4: Multi-seed experiments to explore the stability of our RL-based training.

Metricw/o RL: Seed 42w/o RL: Seed 43w/o RL: Seed 44w/o RL: Mean ± SDRL: Seed 42RL: Seed 43RL: Seed 44RL: Mean ± SD
ASR (%)63.0149.9561.6258.19 ± 7.1799.9599.9496.6798.85 ± 1.89
NPV (%)100.0098.91100.0099.64 ± 0.63100.00100.0094.8198.27 ± 3.00
Distinct-10.13860.14130.13650.1388 ± 0.00240.07000.07400.12670.0902 ± 0.0316
Distinct-20.35520.34280.34330.3471 ± 0.00700.21590.22530.36800.2697 ± 0.0853
Self-BLEU0.66380.67680.67320.6713 ± 0.00670.79580.78530.63100.7374 ± 0.0923
Pos-Std0.06730.07420.07140.0709 ± 0.00350.09620.08830.11750.1007 ± 0.0151
Pos-Entropy0.21560.40280.31140.3100 ± 0.09360.21660.21490.52220.3179 ± 0.1769

To evaluate the stability of our RL-based training framework, we conducted experiments with multiple random seeds (Seed 42, 43, and 44) on the car dataset (Table X). The results demonstrate that incorporating RL significantly mitigates the sensitivity to initialization observed in the non-RL baseline (WrapperSBA w/o RL). First, regarding attack effectiveness (ASR), the non-RL variant exhibits high variance, yielding a mean ASR of only 58.19% with a substantial standard deviation of 7.17% (ranging from 49.95% to 63.01% across seeds). In contrast, the full WrapperSBA model not only boosts the mean ASR to an impressive 98.85% but also drastically reduces the standard deviation to 1.89%, maintaining consistently high attack success (96.67% – 99.95%) across all seeds. This confirms that our RL mechanism substantially improves both the robustness and stability of the attack. Second, in terms of behavioral specificity (NPV), both methods maintain exceptionally high levels (over 98%), indicating that the introduction of RL does not compromise the model's ability to reject non-targeted instructions.

Regarding diversity, an expected trade-off between effectiveness and stealthiness is observed. Wrapper diversity metrics (Distinct-1 and Distinct-2) decrease compared to the non-RL version, accompanied by an increase in Self-BLEU. This suggests that to achieve extreme and stable attack success rates (ASR), the RL policy converges toward highly effective backdoor wrapping patterns, thereby sacrificing some vocabulary diversity. However, regarding position diversity, the RL variant performs better, achieving higher Pos-Std (0.1007) and Pos-Entropy (0.3179), which indicates that the RL model distributes the attack positions more uniformly and randomly within the text. In summary, the multi-seed experiments confirm that our RL framework dramatically enhances attack success and training stability while successfully maintaining behavioral specificity and positional diversity.

AuthorsOpenReview ↗

Response to Weakness 5

Weakness 5: Numerical hyperparameters and reproducibility

We agree that the previous version lacked adequate numerical details. Following this comment, we have added a blue hyperparameter table rather than relying only on the prose description of Algorithm 2. The table reports K=4K=4, a learning rate of 10−610^{-6}, LoRA rank 64, LoRA alpha 16, dropout 0.05, and no LoRA bias. It also reports the initial efficacy weights (+6,−6,+3,−5)(+6,-6,+3,-5), diversity weights (1.0,0.5,1.0)(1.0,0.5,1.0), wglb=−3w_{glb} = -3, τdup=0.85\tau_{dup} = 0.85, τrel=0.55\tau_{rel} = 0.55, τsat=0.95\tau_{sat} = 0.95, [smin,smax]=[0.5,2.0][s_{min}, s_{max}] = [0.5, 2.0], γ=0.6\gamma = 0.6, ζ=0.8\zeta = 0.8, target contributions (0.30,0.15,0.10)(0.30,0.15,0.10), and Rmax=5R_{max} = 5. Batch sizes, KL regularization, clipping, gradient normalization, rollout temperatures, and instruction schedules are also included.

Appendix A.3 now explains the scheduling procedure, while the new table supplies its numerical configuration. We will further align every table entry with the method notation and add explicit entries for wurlw_{url}, λp\lambda_p, and the compute budget, which are not unambiguously represented in the current table. The Reproducibility Statement will refer directly to this consolidated table.

Overall, these comments have helped us distinguish a specific ConfGuard failure regime from general defense evasion. We will narrow the claims and correct the reporting issues, while the new Figure 3, multi-seed experiment, and consolidated hyperparameter table provide additional evidence on the detection mechanism, training variability, and reproducibility.

CategoryParameterActual value
OptimizationEpochs / prompt batch size / responses per prompt K30 / 2 / 4
OptimizationPolicy micro-batch size2 generated sequences
OptimizationOptimizer / learning rateAdamW / 1×10⁻⁶
OptimizationKL coefficient / ratio clipping / gradient norm0.04 / 0.2 / 1.0
OptimizationPrecision / maximum input lengthBF16 / 2,048 tokens
RolloutTop-p / maximum new tokens0.95 / 512
RolloutTemperature0.7 for epochs 1–3; 0.9 thereafter
LoRARank / alpha / dropout / bias64 / 16 / 0.05 / none
Initial efficacy weightsASR success / failure / silence / URL without entity+6 / −6 / +3 / −5
Initial diversity weightsDiversity / position / relevance1.0 / 0.5 / 1.0
Initial penaltiesRepetition / additional URL / collapse / off-topic−1 / −2 / −3 / −3
Quality termsRelevance threshold0.55
Quality termsFluency weight / scale / EMA decay0.2 / 6.0 / 0.99
Quality termsGlobal duplication penalty / similarity threshold−3 / 0.85
Quality termsLength coefficient / threshold−0.5 / 1,000 characters
Quality termsCollapse sentence length / total reward clipping500 characters / [−5, 5]
ASR–NPV schedulingEMA decay / initial α / scale range0.6 / 0.5 / [0.5, 2.0]
ASR–NPV schedulingSaturation threshold / shrinkage range0.95 / [0.3, 1.0]
Contribution balancingDiversity / position / relevance targets0.30 / 0.15 / 0.10
Contribution balancingSmoothing / weight bounds / factor bounds0.8 / [0.1, 5.0] / [0.5, 2.0]
Instruction schedulingInstruction tiersStrong / medium / empty
Instruction schedulingEmpty-instruction probabilityBase 0.1 → 0.9, plus 0.15(gap_NPV − gap_ASR), clipped to [0.05, 0.95]
Instruction schedulingStrong-instruction fraction among non-empty prompts0.7 → 0.3
EvaluationPrompts / seed / temperature / beams600 / 0 / 0.2 / 1
Reviewer_T6zLOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper builds on the Shadow-activated Backdoor Attack (SBA) from BadMLLM. In SBA, a multimodal LLM inserts a malicious string (e.g., a crafted URL) into its response whenever its own output mentions a target entity, so no input trigger is needed. The authors argue that existing SBAs ignore output-side concealment. They always append the same fixed sentence, so output-analysis defenses such as ConfGuard can detect them easily.

To fix this, they split the backdoor carrier (the full malicious sentence) into an invariant goal (the URL) and flexible wrappers (varied phrasing and position around it). This moves the similarity between carriers from the sentence level to the lexical level. Plain SFT on a diverse-wrapper dataset fails, with ASR as low as 46.6%. The model can no longer link the entity to the URL when the surrounding text keeps changing.

The proposed method, WrapperSBA, therefore casts implantation as a constrained multi-objective RL problem and trains with GRPO. The reward combines several elements including attack efficacy, fluency and semantic and positional diversity of wrappers.

Experimental results show improved performance with respect to BadMLLM on three datasets and against LLaVA-7B.

Critical strengths

Here I report what in my opinion are the main strength points:

  1. The problem tackled is well explained and the gap considered is real. Existing backdoor attacks, including SBA, hide the trigger on the input side but always emit the same rigid sentence, which output-analysis defenses can catch.

  2. The reward design is rich and covers several elements needed to tackled the problem including semantic and positional diversity, relevance, fluency and repetition penalties.

  3. The attack is tested against many defenses at different stages: ConfGuard, ToT, Probe, BYE, SRD and PurMM. The ConfGuard analysis is careful, varying the fixed-carrier length relative to the sliding window and explaining why the defense is evaded.

Critical weaknesses

Here I report what in my opinion are the main concerns

  1. One of my main concerns is the presentation of the work. According to the introduction and abstract, a core idea (if I understood correctly) is to model the problem as an RL problem i.e. defining and optimizing an objective whose variables are the parameters of a probability distribution. I think a clearer formalization of the problem is needed, in order to make clear such formalization at the beginning of Section 3. Instead of introducing the optimization target in eq. 3, it would be much clearer to start from the definition of the problem setting and then describe how it is tackled i.e. I believe Section 3 should be reframed.

  2. Moreover, figure 2 is too details with respect to the caption description and make difficult to understand the overall approach.

  3. Algorithm 1 should be explained more step by step before providing the pseudocode. Another limitation is on the experiments.

  4. The core experiments (Table 1) is performed only on LLaVA-7B and 13B and only other model (but used as base model) is Qwen 1.5B (in appendix).

Despite these weaknesses, I believe the work is interesting and propose a useful novel algorithm.

Questions for authors
  1. In Table 2 (if I understood correctly), evasion of ConfGuard holds only when the fixed part is shorter than the window (L < 10). What is the token length of the URLs in the main multimodal experiments, and how does ConfGuard (with the default and a larger window) perform on those outputs? Would simple URL-level filtering or an LLM-based judge detect WrapperSBA outputs?

  2. Which components drive the gains? Can you ablate the instruction-decay curriculum and the individual reward terms (diversity, relevance, fluency)? Is the benefit of the dynamic scheduler also visible in the final Table 1 metrics, and not only in the training curves of one dataset?

4 comments
AuthorsOpenReview ↗

Response to Weakness 1-3

We would like to sincerely thank the reviewer for their valuable comments and suggestions, which help us significantly improve the quality of our paper. All modifications in the new version of the paper are marked in blue. We sincerely hope that our response can address your concerns.

Weakness 1: Optimization of Section 3

We have restructured Section 3 to clearly separate the problem formulation from its optimization procedure. The revised Section 3.1 first defines the multimodal input distribution, the conditional response distribution πθ(y∣x)\pi_\theta(y \mid x), and the policy parameters θ\theta as the optimization variables. It then introduces the goal and wrapper decomposition and explicitly formulates backdoor implantation as a constrained multiobjective problem in Eq. (1). The objective maximizes the expected verifiable reward for activation reliability, wrapper diversity, positional variation, contextual relevance, and response quality, subject to a constraint on the false positive rate. We subsequently define the individual reward components before introducing GRPO as the solver for a clipped surrogate of this objective. This organization clarifies the relationship between the attack requirements, the distribution-level optimization problem, and the RL-based training procedure.

Weakness 2: Optimization of Figure 2

In response to your feedback, we have removed non-essential details to present the figures in a simpler and more direct style.

Weakness 3: Detailed analysis of Algorithm 1

We have expanded Section 3.2.4 to explain Algorithm 1 step by step before presenting the pseudocode. Specifically, we now describe four sequential stages, namely initialization of the training states and instruction prompts, response generation and reward computation, group relative advantage estimation and policy updates, and reward adjustment followed by checkpoint selection and memory bank resetting. These explanations clarify how the instruction schedule, reward terms, GRPO updates, and dynamic scheduler are connected during training.

AuthorsOpenReview ↗

Response to Weakness 4-6

Weakness 4:Insufficient experiments on model architectures We agree that evaluating on a broader range of MLLM architectures is crucial to demonstrate the generalizability of our approach. To address this, we have significantly expanded our core experiments. Beyond LLaVA-7B/13B, we have now included InternVL2.5-4B, Gemma3-4B-IT, and MiniCPM-V-2.6-8B. The results demonstrate that our method is architecture-agnostic and consistently outperforms the baseline (BadMLLM) across different models. As shown in the table below, while BadMLLM suffers from severe diversity collapse (e.g., Distinct-1 of 0.01 and Self-BLEU of 1.00), our method maintains high attack effectiveness (ASR) while significantly improving wrapper and positional diversity. We have incorporated these new experimental results into Table 2 of the revised manuscript.

ModelMethodASR ↑\uparrowTNR ↑\uparrowDistinct-1 ↑\uparrowDistinct-2 ↑\uparrowSelf-BLEU ↓\downarrowStd ↑\uparrowEntropy ↑\uparrow
LLaVA-13BBadMLLM100.00100.000.010.011.000.100.59
Ours (w/o RL)81.1397.870.210.560.500.120.57
Ours (RL)97.83100.000.150.480.490.170.81
InternVL2.5-4BBadMLLM92.25100.000.010.011.000.090.58
Ours (w/o RL)77.2798.890.210.550.470.100.48
Ours (RL)92.9898.840.180.520.510.080.46
Gemma3-4B-ITBadMLLM87.33100.000.010.011.000.070.49
Ours (w/o RL)74.4999.020.210.570.450.120.65
Ours (RL)94.8795.180.180.540.500.130.68
MiniCPM-V-2.6-8BBadMLLM94.4198.250.040.050.980.130.65
Ours (w/o RL)88.0099.000.210.540.570.110.56
Ours (RL)90.27100.000.200.530.590.070.49

Weakness 5:Question about ConfGuard

ConfGuard validates a specific mechanism of wrapper diversification. ConfGuard's detection performance depends strongly on both the target URL length and its sliding-window configuration. We optimize the paper by designing the new Figure 3 to replace the Table 2, reporting "Diverse wrappers benefit the stealthiness of WrapperSBA by shrinking the fixed carrier below ConfGuard’s window". Specifically, let LcL_c denote the length of the complete fixed carrier and Lg<LcL_g<L_c the length of its invariant payload. BadMLLM exposes a shared high-confidence region of length LcL_c, whereas WrapperSBA reduces the shared region to LgL_g and diversifies the remaining wrapper. WrapperSBA obtains 4.8–22.8% TPR for short cores , compared with 100% for fixed-wrapper SBA. Reducing the detection window introduces substantial practical trade-offs. Natural responses frequently contain expressions similar to the invariant payload, making smaller windows increasingly susceptible to false positives. Furthermore, shorter windows substantially increase computational overhead due to the larger number of candidate segments that must be examined. Therefore, although ConfGuard can potentially detect shorter invariant carriers by reducing its window size, doing so comes at the cost of increased false-positive rates and computational complexity. This further demonstrates the practical value of wrapper diversification.

Weakness 6:Ablation Study

Additional effectiveness rewards of RL and diversity rewards serve distinct purposes. Effectiveness rewards of RL recovers attack effectiveness, whereas diversity rewards of RL prevent carrier collapse.

SettingASRNPVD-1 ↑D-2 ↑Self-BLEU ↓Pos-Std ↑Pos-Entropy ↑
Full RL97.50%98.69%0.210.360.580.120.61
w/o diversity rewards99.95%100%0.07020.21250.79870.06450.1026

Without diversity rewards, ASR and NPV remain near saturation, but lexical and positional diversity deteriorate markedly. The model therefore exploits a high-ASR but repetitive-output shortcut unless diversification is explicitly optimized. This directly isolates the independent contribution of the diversity objectives to output-side stealth.

Generation quality is also constrained during optimization through relevance and fluency rewards, together with repetition, multiple-URL, collapse, and excessive-length penalties. NPV further measures behavioral specificity, and Tables 11–13 provide actual generated responses for qualitative inspection. These examples allow readers to directly inspect response relevance, fluency, and overall generation quality, providing complementary evidence beyond aggregate attack-success metrics.

I thank the authors for their detailed response. It resolved my main concerns, and for this reason (and after reading the other reviews and responses), I decided to keep my positive score. I would just like to suggest, for improved clarity, adding a short paragraph before Section 3.2.1 briefly describing the components of WrapperSBA, to better connect the paragraph in Section 3.2 with the following subsections.

AuthorsOpenReview ↗

Thank you for your positive feedback on our method.

Thank you for your positive feedback on our method. Your constructive suggestions are highly valuable for improving our paper. Comments from other reviewers have also made us realize that the presentation of our experiments may cause ambiguity. Given that other reviewers have not provided timely feedback, we are unable to ascertain whether this ambiguity has been fully addressed. Therefore, we have decided to withdraw the manuscript for further refinement.

Reviewer_mar6OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates a better attacking method for MLLM, i.e., WrapperSBA which separates a fixed attack goal from variable wrappers in multimodal backdoor attacks.

It uses SFT followed by multi-objective RL, with adaptive reward weights and instruction decay, to balance output diversity and attack success.

Experiments on three entity datasets with LLaVA-7B/13B report that the RL stage restores the ASR lost under SFT on diverse wrappers, and that WrapperSBA reduces the detection effectiveness of the evaluated defenses relative to BadMLLM.

Critical strengths
  • The paper raises a useful defense question: how much does detection depend on repetitive malicious responses?
  • The distinction between the goal and its wrapper is intuitive and makes the intended behavior easy to follow.
Critical weaknesses
  • Performance gains are uneven. Probe on Sandwich and PurMM on Car support the method (Tables 3, 5), but several defenses already struggle with BadMLLM (Tables 4, 6, 7). ConfGuard evasion depends strongly on goal length (Table 2), while CleanGen remains effective in the unimodal experiments (Table 9). These results warrant a narrower claim about evasion.

  • For SBA, ASR is conditional on entity mentions (Eq. 1), but mention rates do not seem to be reported. The models could be scored on different subsets of responses, making the comparison harder to interpret.

  • Tables 1 and 10 show that RL recovers ASR after diversification, but do not isolate the contribution of its diversity objectives from additional training. An otherwise matched RL control without these objectives is missing, as is an independent clean-task quality evaluation.

  • Some statements about defense performance seem inconsistent with the reported numbers. On p. 9 (lines 462–463), ToT is described as "achieving higher TPRs on the fixed-wrapper variants," yet Figure 5 and Table 4 give TPRs of 9.8% for fixed-wrapper BadMLLM and 95.8% for WrapperSBA-RL on Sandwich. Lines 461–462 report SRD TPR/FPR as 14.3%/10.6%; Figure 5 and the Sandwich RL row of Table 7 give 7.0%/12.5%.

Questions for authors
  1. What are the token lengths and ConfGuard results for the main-experiment URLs listed in Appendix A.1?
  2. Which reward, instruction, and scheduling choices actually matter? Please isolate their effects and report the key hyperparameter values.
3 comments
AuthorsOpenReview ↗

Response to Concern 1

We would like to sincerely thank the reviewer for their valuable comments and suggestions, which help us significantly improve the quality of our paper. All modifications in the new version of the paper are marked in blue. We sincerely hope that our response can address your concerns.

Concern 1 — Scope and interpretation of the evasion results

The limited effectiveness of existing defenses against BadMLLM does not indicate a weakness of our method. Our experiments demonstrate that WrapperSBA further improves output-side stealth against BadMLLM while preserving the attack effectiveness against existing defenses.

BadMLLM already provides a strong baseline. Even when employing the latest state-of-the-art (SOTA) defenses, only a limited number of experimental results can explicitly highlight the significant degradation in defense effectiveness caused by diverse wrappers. The inherent limitations of current defenses restrict the ability to fully demonstrate the experimental outcomes of our work. We demonstrate this limitation at the section Conclusions & Limitations .The defenses evaluated in our experiments are representative state-of-the-art methods proposed in 2025 and 2026. Their limited effectiveness against BadMLLM therefore reflects the inherent difficulty of defending against this attack paradigm, rather than a limitation specific to WrapperSBA.**. In addition, we additionally evaluated BadNets, BadNL, Sleeper-Agents, LLMBkd, Homoglyph, Combined, and UAT attacks to address this limitation.

ConfGuard validates a specific mechanism of wrapper diversification. ConfGuard's detection performance depends strongly on both the target URL length and its sliding-window configuration. We optimize the paper by designing the new Figure 3 to replace the Table 2, reporting "Diverse wrappers benefit the stealthiness of WrapperSBA by shrinking the fixed carrier below ConfGuard’s window". Specifically, let LcL_c denote the length of the complete fixed carrier and Lg<LcL_g<L_c the length of its invariant payload. BadMLLM exposes a shared high-confidence region of length LcL_c, whereas WrapperSBA reduces the shared region to LgL_g and diversifies the remaining wrapper. WrapperSBA obtains 4.8–22.8% TPR for short cores , compared with 100% for fixed-wrapper SBA. Reducing the detection window introduces substantial practical trade-offs. Natural responses frequently contain expressions similar to the invariant payload, making smaller windows increasingly susceptible to false positives. Furthermore, shorter windows substantially increase computational overhead due to the larger number of candidate segments that must be examined. Therefore, although ConfGuard can potentially detect shorter invariant carriers by reducing its window size, doing so comes at the cost of increased false-positive rates and computational complexity. This further demonstrates the practical value of wrapper diversification.

Clarification of the Role of Output-Side Stealth in Table 9. The results in Table 9 reinforce the importance of jointly considering input-side and output-side stealth. For BadNets, Combined, and UAT, their input triggers can already be detected by CleanGen due to their insufficient input-side stealth. This does not contradict the importance of output-side stealth. Our approach is better understood as a complementary enhancement to existing backdoor attack paradigms, rather than a replacement for input-side stealth mechanisms. The ability of a backdoor attack to evade defenses depends on the joint stealthiness of its input and output behaviors. Table 9 is intended to illustrate this potential single-point-of-failure issue: an attack can remain vulnerable if either its input trigger or its output carrier is detectable, and robust defense evasion requires both sides to be sufficiently concealed. Therefore, the results in Table 9 reinforce the importance of jointly considering input-side and output-side stealth.

AuthorsOpenReview ↗

Response to Concern 2-4

Concern 2 — Entity mention rates and evaluation subsets

The evaluation set is fixed before response generation, and no test samples are selected or filtered based on model outputs. Specifically, for each target entity, we construct target_qa.json and non_target_qa.json with identical numbers of samples. Consequently, the entity mention rate across the combined evaluation set is exactly 50% by construction. Furthermore, all models are evaluated using identical input prompts, question ordering, instructions, and decoding parameters. Therefore, the evaluation does not involve model-dependent sample selection, and all models are assessed on the same predefined test set. We will clarify this evaluation protocol in the revised manuscript to avoid potential misunderstandings regarding the comparability of ASR measurements.

Concern 3 — Diversity objectives and generation quality

Additional effectiveness rewards of RL and diversity rewards serve distinct purposes. Effectiveness rewards of RL recovers attack effectiveness, whereas diversity rewards of RL prevent carrier collapse.

SettingASRNPVD-1 ↑D-2 ↑Self-BLEU ↓Pos-Std ↑Pos-Entropy ↑
Full RL97.50%98.69%0.210.360.580.120.61
w/o diversity rewards99.95%100%0.07020.21250.79870.06450.1026

Without diversity rewards, ASR and NPV remain near saturation, but lexical and positional diversity deteriorate markedly. The model therefore exploits a high-ASR but repetitive-output shortcut unless diversification is explicitly optimized. This directly isolates the independent contribution of the diversity objectives to output-side stealth.

Generation quality is also constrained during optimization through relevance and fluency rewards, together with repetition, multiple-URL, collapse, and excessive-length penalties. NPV further measures behavioral specificity. In addition to these reward-based constraints, we provide extensive qualitative evidence in Tables 11–13, spanning nearly three pages, which presents actual model-generated responses. These examples allow readers to directly inspect response relevance, fluency, and overall generation quality, providing complementary evidence beyond aggregate attack-success metrics.

Concern 4 — Numerical and textual inconsistencies

A defense is considered effective only when it achieves both a high true positive rate (TPR) and a low false positive rate (FPR).Our intended interpretation is that, compared with the fixed-wrapper BadMLLM variant, WrapperSBA generally induces higher TPRs but also substantially higher FPRs. This distinction is important because the increased TPR does not necessarily indicate improved detectability in practice, as it comes at the cost of misclassifying more benign samples as malicious.Specifically, BadMLLM exhibits relatively low TPR and low FPR, whereas WrapperSBA tends to induce higher TPR and higher FPR. This suggests that WrapperSBA makes existing detection methods more likely to flag benign responses, thereby preventing them from simultaneously achieving effective attack detection and an acceptable false alarm rate.

AuthorsOpenReview ↗

Response to Concern 5-6

Concern 5 — Main-experiment URL lengths and ConfGuard Using the LLaVA-v1.5-7B LlamaTokenizer with add_special_tokens=False, the URL lengths are:

GoalURL token length
Car20
Sandwich27
Laptop27

ConfGuard's detection performance depends strongly on both the target URL length and its sliding-window configuration. We optimize the paper by designing the new Figure 3 to replace the Table 2, reporting "Diverse wrappers benefit the stealthiness of WrapperSBA by shrinking the fixed carrier below ConfGuard’s window".

Concern 6 — Reward, instruction, scheduling, and hyperparameters

Our central claims do not require every one of the 12 reward terms to be independently necessary. They require the key design choices to have identifiable effects, which we isolate through matched, hypothesis-driven controls. Beyond wrapper diversification, another key component of our method is an adaptive reward-adjustment mechanism designed to reduce sensitivity to manually specified initial hyperparameters. Specifically, the mechanism dynamically adjusts reward coefficients according to the learning progress of individual reward components, helping maintain a balanced optimization process throughout RL training. We list the details of the related hayperparameters in Appendix D.

We acknowledge the reviewer's interest in isolating the contributions of individual reward components. The reward components serve complementary purposes rather than acting as interchangeable optimization objectives. Their functional contributions are motivated by the specific behavioral constraints they are designed to enforce. However, our reward function contains 12 components, resulting in 212=4,0962^{12}=4,096 possible inclusion/exclusion configurations. Exhaustively evaluating all combinations would introduce substantial computational overhead without necessarily providing proportionate insight.

Instead, we specify the function of each reward component and evaluate the overall framework through our existing ablation experiments. We believe this provides a practical characterization of the reward design without requiring an exhaustive combinatorial search.

CategoryParameterActual value
OptimizationEpochs / prompt batch size / responses per prompt K30 / 2 / 4
OptimizationPolicy micro-batch size2 generated sequences
OptimizationOptimizer / learning rateAdamW / 1×10⁻⁶
OptimizationKL coefficient / ratio clipping / gradient norm0.04 / 0.2 / 1.0
OptimizationPrecision / maximum input lengthBF16 / 2,048 tokens
RolloutTop-p / maximum new tokens0.95 / 512
RolloutTemperature0.7 for epochs 1–3; 0.9 thereafter
LoRARank / alpha / dropout / bias64 / 16 / 0.05 / none
Initial efficacy weightsASR success / failure / silence / URL without entity+6 / −6 / +3 / −5
Initial diversity weightsDiversity / position / relevance1.0 / 0.5 / 1.0
Initial penaltiesRepetition / additional URL / collapse / off-topic−1 / −2 / −3 / −3
Quality termsRelevance threshold0.55
Quality termsFluency weight / scale / EMA decay0.2 / 6.0 / 0.99
Quality termsGlobal duplication penalty / similarity threshold−3 / 0.85
Quality termsLength coefficient / threshold−0.5 / 1,000 characters
Quality termsCollapse sentence length / total reward clipping500 characters / [−5, 5]
ASR–NPV schedulingEMA decay / initial α / scale range0.6 / 0.5 / [0.5, 2.0]
ASR–NPV schedulingSaturation threshold / shrinkage range0.95 / [0.3, 1.0]
Contribution balancingDiversity / position / relevance targets0.30 / 0.15 / 0.10
Contribution balancingSmoothing / weight bounds / factor bounds0.8 / [0.1, 5.0] / [0.5, 2.0]
Instruction schedulingInstruction tiersStrong / medium / empty
Instruction schedulingEmpty-instruction probabilityBase 0.1 → 0.9, plus 0.15(gap_NPV − gap_ASR), clipped to [0.05, 0.95]
Instruction schedulingStrong-instruction fraction among non-empty prompts0.7 → 0.3
EvaluationPrompts / seed / temperature / beams600 / 0 / 0.2 / 1

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

Related papers

Discussion (0)

Sign in to comment.