acceptodds
Withdrawn from ICLR 2027

A Systematic Study of Cross-Modal Typographic Attacks on Audio-Visual Reasoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

As audio-visual multi-modal large language models (MLLMs) are increasingly deployed in safety-critical applications, understanding their cross-modal vulnerabilities is essential. To this end, we introduce Multi-Modal Typography, a systematic study examining how perturbations added to speech and on-screen text impact audio-visual reasoning of closed- and open-source models. For instance, on the WorldSense benchmark, the attack success rate (ASR) against Gemini 3.7 Flash increased from 9.53% to 43.41%, highlighting its vulnerability. We analyze the cross-modal effects of unimodal and multimodal attacks and evaluate how specific properties of the attacks impact the overall multi-modal reasoning. By applying fine-tuning as a mitigation, we lower the ASR of Qwen2.5-Omni-7B on WorldSense from 60.6% to 52.0%. Our findings across multi-modal reasoning tasks, content moderation, and offline robot task selection benchmarks establish multi-modal typography as a critical and underexplored attack strategy in multi-modal reasoning. Code and data will be made publicly available.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
multimodal large language models, audio-visual models, multimodal robustness, multi-modal typography, spoken semantic injection, multimodal evaluation, safety evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_oESrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Multi-Modal Typography, a systematic study of how misleading speech and on-screen text affect audio-visual reasoning in MLLMs. It examines the cross-modal effects of unimodal and multimodal attacks, as well as the influence of attack properties. The experiments cover audio-visual reasoning, content moderation, and offline robot task selection, and evaluate prompting and fine-tuning as mitigation methods.

Critical strengths

The experiments are comprehensive, covering different attack modalities and their combinations, as well as several mitigation methods.

Critical weaknesses
  1. The motivation is not sufficiently clear given the existing work on typographic attacks and audio injection attacks. The paper should explain more clearly what questions remain unanswered and how its study addresses them.

  2. The attack setting of the main experiments assumes knowledge of the question and answer space and directly suggests a wrong answer, while random speech and noise barely affect accuracy in Table 4a. The practical relevance of this setting is unclear when the attacker does not know the user’s question or answer choices. Moreover, while the paper analyzes answer scores and attention, it remains unclear whether explicit reasoning can reduce the attacks’ effectiveness. Comparing direct answering with chain-of-thought reasoning before giving the final answer would help clarify this.

Questions for authors

Please see the weaknesses

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

The paper studies whether misleading synthesized speech and on-screen text can redirect audio-visual models toward incorrect answers. It compares individual and combined attacks across three benchmarks, examines speech properties and human detectability, and evaluates content moderation, offline robot task selection, and fine-tuning as a defense.

Critical strengths

The random-noise and unrelated-speech controls provide useful evidence that targeted semantic content contributes to the failures. Evaluating visually grounded questions also demonstrates that misleading audio can affect predictions even when the visual input remains unchanged. This gives the study some empirical value.

Critical weaknesses

Insufficiently controlled modality comparisons. Tables 2–3 keep the attack target fixed, but the paper does not adequately specify or match wording, repetition, and exposure across speech, on-screen text, and prompt injections. Visual-overlay settings are also insufficiently documented. Consequently, differences in attack success cannot be confidently attributed to modality alone, weakening a central contribution and its reproducibility.

Key conclusions exceed the supporting validation. Appendix A.7 explicitly states that human detectability and model attack success were measured on different videos. This does not establish that attacks fooling the model were imperceptible to humans. Moreover, the safety and defense experiments rely on Qwen2.5-Omni-7B, despite substantial model-dependent differences in the main results. These gaps leave the stealth and broader practical conclusions insufficiently supported.

Questions for authors

Clarification: What exact wording, repetition, duration, and visual-overlay settings were used in Tables 2–3? How were differences in attack exposure controlled?

Additional validation: Can you report attack success on the same clips and speech tracks judged undetectable by human participants, with uncertainty estimates?

Additional validation: Do the safety and defense findings hold for the stronger models already included in Table 1? This would materially affect confidence in their generality.

Reviewer_wKT4OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies targeted semantic injection into audio-visual MLLMs through synthesized speech, on-screen text, and text prompts. It evaluates single-modality attacks and aligned or conflicting audio-visual attacks on several QA benchmarks, analyzes attack parameters and human detectability, and includes content-moderation and offline robot-task-selection experiments. Adversarial fine-tuning provides a modest reduction in attack success on Qwen2.5-Omni-7B.

Critical strengths
  1. The evaluation covers multiple audio-visual models and modality-specific QA tasks.
  2. Target-specific ASR and random-speech controls help distinguish targeted attacks from general degradation.
Critical weaknesses
  1. The main experiments inject question-specific wrong answers on general QA benchmarks such as WorldSense, but the practical relevance of this setup remains insufficiently validated.
  2. Safety evaluation covers only MetaHarm and I2P and is limited to Qwen2.5-Omni, without testing the newer models.
  3. Adversarial fine-tuning reduces ASR from 60.6% to 52.0%, leaving the model highly vulnerable and demonstrating limited mitigation effectiveness.
  4. The paper insufficiently explains the performance differences across injection modalities, leaving the roles of cue recognition and modality fusion unclear.
Questions for authors
  1. Could you analyze the sources of performance differences across injection modalities in more detail, e.g., unequal understanding of speech and text, or greater reliance on text during fusion despite understanding both?
  2. Could you evaluate your attacks on safety benchmarks such as MetaHarm and I2P using frontier models?
  3. Could you evaluate whether more adversarial training data or additional fine-tuning steps further improve robustness?
Reviewer_YFE1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper examines whether the typographic vulnerability of vision-language models extends to spoken cues in audio-visual MLLMs, and whether models respond consistently to the same misleading cue across text prompts, speech and on-screen text. Across eight models and three benchmarks, it tests speech naming an incorrect answer, alone or alongside visual and prompt cues, with additional human-perception studies, downstream evaluations and LoRA-based mitigation. However, the strongest effects require loud, repeated speech explicitly stating the answer, cross-modal consistency is assessed only descriptively, and ASR remains at 52.0% after mitigation.

Critical strengths

The cross-modal finding holds across a broad set of models and tasks. Using benchmarks that split questions by modality, the paper shows that speech injected into an unchanged video hurts visually grounded questions.

Critical weaknesses
  • The headline vulnerability depends on an explicit, loud, question-specific cue, making the claim of stealth difficult to justify.

  • The paper attributes PandaGPT’s near-zero ASR to poor speech recognition rather than robustness (Section 4.2), but does not measure whether models can understand the injected phrases, for example through transcription accuracy. It is therefore unclear whether lower attack success reflects stronger safety or simply poorer speech perception.

  • The novelty is limited. How does audio typography differ, conceptually and empirically, from prior audio prompt-injection attacks (Bagdasaryan et al., 2023; Chen et al., 2026) and misleading-audio studies (Hou et al., 2025)? What finding is new beyond the observation that models follow explicit answer cues regardless of channel?

Questions for authors

How does susceptibility relate to each model’s ability to transcribe the injected phrase? Can you report transcription accuracy per model?

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

Related papers

Discussion (0)

Sign in to comment.