acceptodds
Under review as a conference paper at ICLR 2027

FusionBreak: Attacking Multimodal Fusion While Preserving Unimodal Decisions

Abstract

Multimodal models combine information from multiple modalities, but correct unimodal predictions do not necessarily guarantee a correct fused prediction. Conventional adversarial evaluations typically focus on whether the final multimodal prediction becomes incorrect without distinguishing errors that emerge at the fused output from those already present in individual modalities. We study a stricter condition in which the unimodal classifiers remain correct while the fused model becomes incorrect. We introduce FusionBreak, an adversarial attack that preserves unimodal decisions while causing the fused model to misclassify, and define the Fusion-Only Success Rate (FOSR) to measure this behavior. Experiments across two datasets and four fusion schemes show that fusion-only failures occur across different fusion designs. In contrast, standard PGD can achieve a high fused attack success rate with 0% FOSR, showing that conventional attack success does not distinguish whether the fused model fails while the unimodal predictions remain correct. Ablation experiments further demonstrate the importance of preserving unimodal decisions. We also evaluate a detector-aware variant of FusionBreak that adds detector-evasion terms to the attack objective. This reduces detection by several generic adversarial detectors while largely preserving the effectiveness of the attack.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Adversarial attacks, multimodal learning, multimodal fusion, adversarial robustness, audio-visual learning, fusion robustness, adversarial examples, multimodal security

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_vf8oOpenReview ↗
Rating
2(+1)
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces FusionBreak to focus on attacking multimodal models while preserving correct unimodal predictions. They also introduce a detector-aware loss to further reduce the detection rate. Experiments on two audio-visual datasets demonstrate the effectiveness of the proposed method.

Critical strengths
  1. This paper is well-written, I can quickly understand its research objective. The research question is also interesting: the authors try to attack multimodal fusion while preserving correct unimodal predictions, thus hiding the attack at the unimodal level.

  2. Proposed method can reduce the detection ratio under different detectors.

Critical weaknesses
  1. From my perspective, the main contribution of this paper is introducing a new attack setting and modifying the loss function of PGD rather than proposing a novel algorithm. Meanwhile, I'm not sure about the practical importance of the proposed setting.

  2. The paper only compares the proposed method with standard PGD, without including other adversarial attack methods. Beyond the FOSR metric, Appendix D shows that FusionBreak can also exhibit lower detection rates than PGD without the detector-aware loss. Thus, I think this potential to evade detectors could be an important contribution of the method. However, this also needs more comparisons with other baselines considering evading detection, e.g., PGD with detector-aware loss.

  3. Datasets and targeted model are limited. (2 datasets and CNN)

Questions for authors
  1. Can the authors provide some practical scenarios or applications where the proposed fusion-only attack would be particularly relevant? This is important for me to estimate the value of proposed method.

  2. Can the proposed method generalize to more advanced pretrained multimodal models and other datasets?

8 comments
AuthorsOpenReview ↗

Clarification of Contributions, Practical Importance, and Expanded Evaluation

We respectfully disagree that a clear rejection adequately reflects the scientific contributions and evidence presented in our work, particularly given the positive aspects acknowledged in the review.

1. Practical importance

The failure pattern studied in our paper has been observed in previous research.

For example, Lingenfelser et al. (ICMI 2011, DOI: 10.1145/2070481.2070487) investigated multimodal emotion recognition and reported cases where both audio and video classifiers made correct predictions, but the fused classifier produced an incorrect result.

Our work investigates whether an adversary can deliberately induce this type of failure.

The broader security importance of multimodal fusion has also been demonstrated in autonomous driving. Hallyburton et al. (USENIX Security 2022) studied adversarial attacks against camera-LiDAR fusion systems. Although their setting differs from ours, it illustrates the practical consequences of vulnerabilities in multimodal decision-making.

Our original experiments achieve FOSR values of 77.09% on RAVDESS and 97.67% on CREMA-D, compared with 0% for standard PGD. This demonstrates why conventional attack success rates alone do not capture the failure condition we study.

2. Methodological contribution

We respectfully disagree that our contribution is simply a modification of the PGD loss.

Our original work introduces a constrained adversarial objective, decision-aware candidate selection, and an evaluation metric that explicitly requires correct unimodal predictions.

We further investigated this problem through an additional paired perturbation rescaling procedure and strict interaction checks.

The procedure searches combinations of audio and video perturbation strengths. A strict interaction success occurs only when the individual perturbations and their additive reference preserve the fused decision, while their joint application causes an incorrect prediction.

This allows us to distinguish joint interaction failures from ordinary disagreement between classifiers.

3. Additional baselines

We expanded the experimental evaluation to an 18-method portfolio, including conventional PGD, feasible PGD, preservation-guided search, strict PGD with different optimization budgets, single-modality attacks, and component ablations.

For example, on AVE cross-attention, the 20-step strict PGD baseline achieves a strict interaction success rate of 65.7%, compared with 91.9% for the combined rescaling and direct-search procedure.

We also evaluated an 80-step PGD baseline and measured computational costs. These comparisons distinguish the contribution of the additional search procedure from simply increasing gradient optimization.

Detector evasion is a complementary analysis, not the central contribution of our paper.

4. Stronger models and datasets

We expanded the experiments from two to three datasets and evaluated four fusion architectures using pretrained wav2vec 2.0, DeiT, AST, and CLIP encoders.

On AVE, the evaluated cross-attention models achieve clean fused balanced accuracies of 93.1% and 94.8%. The combined attack achieves approximately 91.9% strict interaction success on both seeds.

These experiments demonstrate the studied failure condition in stronger pretrained audio-visual systems.

5. Overall assessment

The review acknowledges the importance of our research question and the effectiveness of our original method against several generic detectors. Together with the methodological contributions and additional experimental evidence, we believe these findings deserve a more balanced assessment.

Regardless of whether the reviewer changes their rating, we maintain that a score of 1 does not fairly reflect the scientific contribution and supporting evidence of our work.

Reviewer_vf8oOpenReview ↗

Many thanks for your rebuttal and clarification

I agree that attack fusion embedding is important. But my major concern is: why do we need to attack the fusion embedding while not changing the prediction in a single modality? So I say understanding this setting is really important for me. Based on this, I say that "Detector evasion" may be a potential reason?

For the limited technical contribution: I think my review is not suitable. I check some classical papers, I think one of their core contribution is also to propose an important setting and design proper training methods (e.g., VLAttack or Adversarial Prompt Tuning). But here, "important setting" is critical, for example, VLAttack is useful because they revealed the blind spot that more modalities naturally introduce more vulnerabilities.

If possible, can u give me a brief introduction to your motivation of your proposed settings? I will increase my rating to 2 (based on the current rebuttal). Indeed, I think your setting is interesting, but I still maintain the negative assessment before understanding the applications or motivation. Thanks again for your rebuttal.

Reviewer_vf8oOpenReview ↗

Regarding the 3. Additional baselines and 4. Stronger models and datasets

Can u include these results in your current submission?

Reviewer_S2E1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates scenarios where unimodal classifiers remain correct while a fused model becomes incorrect. This paper introduce FusionBreak—a method for crafting adversarial attacks that cause a fused model to misclassify an input while leaving the decisions of the individual unimodal classifiers unchanged—and define the Fusion-Only Success Rate (FOSR) to quantify this phenomenon.

Critical strengths

The paper investigates a specific multimodal adversarial failure scenario in which unimodal classifiers remain correct while the fused model produces incorrect predictions. This provides an alternative perspective for evaluating multimodal robustness beyond conventional attack success rates.

Critical weaknesses

W1: Although the paper identifies a specific type of multimodal adversarial failure and proposes the FusionBreak attack objective and optimization method, the proposed approach essentially combines and adapts existing techniques without introducing a fundamentally novel attack mechanism or optimization strategy. The primary novelty lies in redefining the attack success criterion by imposing additional constraints to preserve unimodal prediction correctness within an existing multimodal adversarial attack framework, rather than making a substantial algorithmic advancement. More importantly, since the paper positions this fusion-specific adversarial failure as its central research problem, a deeper investigation into its underlying causes and mechanisms is necessary. However, the experiments primarily focus on attack success rates, without systematically analyzing changes in unimodal representations, confidence distributions, or the vulnerabilities of different fusion mechanisms. Consequently, the paper fails to provide sufficient insight into the nature of these failures. Overall, the limited methodological novelty and lack of in-depth technical analysis prevent the paper from offering a sufficiently comprehensive understanding of its central research problem, making the claimed contributions difficult to justify.

W2: FusionBreak only ensures that unimodal prediction labels remain correct after the attack, which does not imply that the underlying unimodal representations or information remain intact. This is a critical issue that is not adequately addressed in the paper. The current experiments demonstrate that the fused model can produce incorrect predictions even when the unimodal classifiers remain correct, but they fail to distinguish whether these errors arise from vulnerabilities intrinsic to the fusion mechanism or from perturbations to the underlying unimodal representations. Therefore, the empirical evidence is insufficient to fully substantiate the paper's central claims regarding fusion-specific vulnerabilities. Furthermore, the evaluation is restricted to relatively simple CNN-based encoders and four conventional fusion schemes, without experiments on MLLMs or more sophisticated Transformer-based fusion architectures. This substantially limits the representativeness of the experimental evaluation and the generalizability of the findings. In addition, the paper suffers from poor organization, further undermining its overall coherence and persuasiveness.

Questions for authors
  1. FusionBreak primarily combines existing gradient-based optimization techniques, unimodal prediction preservation constraints, and candidate selection strategies. What are its key differences and technical advantages compared to existing constrained adversarial optimization methods?

  2. Preserving unimodal prediction labels does not necessarily imply that unimodal representations and confidence distributions remain stable. How do the authors distinguish whether the observed fusion failures arise from vulnerabilities intrinsic to the fusion mechanism or from perturbations to unimodal representations? Could the authors conduct causal intervention experiments or other systematic analyses to further investigate the underlying mechanisms of these failures?

  3. Could the authors further evaluate the effectiveness of FusionBreak on stronger pretrained multimodal models or more sophisticated Transformer-based fusion architectures to demonstrate the generalizability of the proposed attack?

1 comment
AuthorsOpenReview ↗

Clarification of Methodological Contributions and Additional Experimental Evidence

We appreciate the reviewer's comments regarding methodological novelty, the causes of fusion failures, and experimental evaluation. We have conducted additional experiments and extended the original attack procedure while retaining the same central research question.

1. Methodological contribution

The original contribution is to explicitly construct adversarial examples that preserve correct unimodal decisions while changing the fused decision.

To strengthen the optimization approach, we developed an additional paired perturbation rescaling procedure.

Starting from gradient-generated perturbations, the procedure independently adjusts their strengths in audio and video. It evaluates 33 strengths per modality, producing 1,089 candidate combinations. By caching the unimodal representations, these combinations require only 66 distinct branch-encoding evaluations.

Unlike taking additional gradient steps, this procedure searches combinations of perturbation strengths along the generated directions.

We also combine this search with direct optimization under stricter interaction constraints.

The following results compare these approaches using the same strict interaction success criterion:

DatasetStrict PGD-20Rescaling searchCombined search
RAVDESS8.2%15.0%19.6%
CREMA-D19.8%11.9%25.6%
AVE (cross-attention)65.7%89.6%91.9%

The combined procedure includes both searches. We also evaluated an 80-step PGD baseline and measured the computational costs of the different methods.

2. Understanding joint fusion failures

We agree that preserving unimodal labels does not imply that their internal representations remain unchanged. Our objective is to identify failures that occur despite correct unimodal decisions.

To investigate the role of interactions more directly, we added single-modality substitution tests and an additive reference.

A strict success requires the fused classifier to remain correct when either perturbation is applied alone and when their individual logit changes are added, but to become incorrect when both perturbations are applied together.

We also require unimodal prediction margins of at least 0.05 and bound their probability-distribution changes using KL divergence.

These checks provide a more detailed characterization of the failure than correct unimodal labels alone.

3. Stronger models and broader evaluation

We expanded the evaluation to three datasets: RAVDESS, CREMA-D, and AVE. The additional experiments use pretrained wav2vec 2.0, DeiT, AST, and CLIP encoders.

We also evaluate pooled Transformer, cross-attention, gated residual, and bilinear fusion architectures.

For example, on AVE, the two evaluated cross-attention models achieve clean fused balanced accuracies of 93.1% and 94.8%. The combined attack achieves a strict interaction success rate of approximately 91.9% on both seeds.

We additionally evaluated multiple optimization baselines, component ablations, and computational costs.

These results strengthen both the methodological and experimental contributions and provide more detailed evidence about the joint interaction failures that motivate our work.

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

The paper proposes FusionBreak, a white-box attack on audio-visual classifiers that perturbs both modalities so that the fused model misclassifies while the audio and video classifiers still predict the correct label. It defines the Fusion-Only Success Rate (FOSR) over "eligible" samples on which all three clean models are correct. The attack uses multi-restart PGD on a fused cross-entropy and margin objective, with a hinge term on unimodal margins and a KL term to the clean unimodal outputs. It returns the first iterate that satisfies the fusion-only condition. On RAVDESS and CREMA-D with four small fusion heads, FusionBreak reaches 77% and 98% FOSR, while standard PGD reaches 0%. A detector-aware variant lowers detection by MSP, Mahalanobis, and feature squeezing, and a unimodal-vs-fused consistency rule detects every successful example by construction.

Critical strengths

Separating errors that originate inside a modality from errors that appear only after fusion is a reasonable question, and FOSR gives a concrete way to ask it. The ablations are clear about where the gain comes from: removing the preservation term or the decision-aware selection sharply lowers FOSR, while the fusion-margin and KL terms matter little.

The paper is upfront that "fusion-only" describes an observed decision pattern, not causal responsibility of the fusion operator. It also reports the consistency check honestly, even though that check undermines the stealth argument.

Critical weaknesses
  1. The headline result is largely true by construction, and the experiments do not show that it is specific to fusion. Standard PGD optimizes only the fused loss, so it is expected to also flip unimodal predictions, and its 0% FOSR carries no information about fusion. The existence of fusion-only examples is also expected. The fused model and the unimodal classifiers are different functions (on CREMA-D the backbones are even fine-tuned jointly with the fusion head), and the budget is large enough to saturate fused attack success at half the default on CREMA-D. With that much room, one decision boundary can be crossed while the others are not. The natural control is missing: replace the fused model with a non-fusion classifier (e.g., an independently trained audio-visual model, a logit-averaging late fusion, or a second unimodal model) and measure the same rate. If that is comparable, the finding is generic disagreement between non-robust models rather than a fusion vulnerability. The video-only attack reaching about 98% FOSR on CREMA-D, with audio untouched, also suggests a fused head that over-relies on one modality. Results are nearly identical across the four fusion schemes, so the paper offers no insight into which fusion mechanisms matter or why. The contribution reduces to a metric and a penalty term.

  2. The threat model and the stealth argument do not hold up. "Unimodal predictions remain correct" requires ground-truth labels, which a deployed monitor does not have. What a monitor can observe is agreement among predictions. Under that view, the stealthy goal is consistent errors across all heads, which ordinary attacks already produce. FusionBreak instead creates the pattern that a trivial consistency check flags: unimodal predictions that agree while the fused output differs. The adaptive evaluation targets generic single-model detectors, which are not the relevant defense, and the consistency defense cannot be evaded within this threat model. The "preservation" is also hard-label with a very small margin (ρ=0.05\rho=0.05), evaluated on the same white-box models being attacked, so unimodal confidence and features are likely heavily altered. A differently trained or ensembled verifier would plausibly disagree. Finally, perturbations are applied in processed spectrogram and frame space, not raw audio and video, and the models are small CNNs with low clean fused accuracy and a small eligible set. The conclusions about multimodal systems in general are not supported at this scale.

Questions for authors
  1. Is the effect specific to fusion? If the fused model is replaced by another classifier over the same inputs (an independently trained audio-visual model, or a late-fusion average), what FOSR does the same attack reach? Without this, the paper cannot separate a fusion-specific weakness from the general fact that two different networks can be pushed apart inside a large perturbation ball.

  2. What does the preservation constraint correspond to at deployment? Nobody observes correctness without labels, so who verifies the unimodal predictions? Would the attack still succeed if the verifiers were different models from those attacked (transfer), or if preservation required a larger confidence margin than 0.05? More generally, an attacker who wants to evade the consistency check must give up exactly the property that defines FusionBreak. How should the contribution be read in that light?

  3. What explains the vulnerability mechanistically? Do the successful perturbations exploit differences between the fused head and the unimodal heads in feature space, or one modality dominating the fused decision? Can a consistency-aware or preservation-aware adversarial training objective remove the failure mode? Does the phenomenon persist with pretrained transformer encoders and with attacks applied in the raw input domain?

1 comment
AuthorsOpenReview ↗

Additional Evidence on Fusion-Specific Failures and Stronger Models

We appreciate the reviewer's detailed comments and recognition of the importance of distinguishing fusion-only failures from conventional adversarial failures. We have conducted additional experiments to address the main concerns.

1. Are the failures specific to multimodal fusion?

We agree that correct unimodal predictions alone do not establish that a failure depends on interactions between modalities. We therefore conducted additional experiments using a stricter interaction criterion.

For each candidate, we evaluate four fused outputs: the clean pair, audio-only perturbation, video-only perturbation, and joint perturbation. We also construct an additive reference from the individual changes:

zadd=z10+z01−z00.z_{\mathrm{add}}=z_{10}+z_{01}-z_{00}.

A strict interaction success requires the unimodal classifiers, both single-substitution fused predictions, and the additive reference to remain correct, while the joint perturbation causes misclassification.

This directly tests whether the joint effect changes a decision that remains correct under the individual changes and their additive combination. In particular, an additive fusion model at the logit level cannot satisfy this strict criterion.

Our additional experiments show:

DatasetPreservation search: FOSRPreservation search: strict successCombined search: strict success
RAVDESS100.0%0.0%19.6%
CREMA-D99.7%0.1%25.6%
AVE (cross-attention)100.0%0.0%91.9%

These results demonstrate why preserving unimodal decisions alone is insufficient for identifying joint interaction failures. The additional criterion measures a more specific condition.

2. Differences between fusion architectures

We expanded the experiments to pooled Transformer, cross-attention, gated residual, and bilinear fusion.

On AVE (seed 43), their strict interaction success rates are 89.9%, 91.9%, 38.9%, and 90.9%, respectively. These differences demonstrate that fusion architecture affects the observed vulnerability.

3. Monitoring and transfer

Ground-truth labels define attack success during evaluation; we do not assume that a deployed monitor has access to them.

We also distinguish decision preservation from detector evasion. A consistency monitor can identify the disagreement pattern targeted by our attack.

To investigate transfer, we evaluated successful candidates on independently trained classification heads with the same frozen encoders. On CREMA-D, 146 of 422 eligible transferred candidates from the combined search retained fusion-only success.

4. Stronger models and input verification

We added AVE to the two original emotion-recognition datasets and evaluated pretrained wav2vec 2.0, DeiT, AST, and CLIP encoders.

The additional attacks operate on waveform and RGB inputs. All 10,965 recorded strict successes across the full evaluation passed verification after saving and reloading the modified WAV and PNG inputs.

We believe these additional analyses provide substantially stronger evidence for the original research question and address the central concerns raised in the review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.