acceptodds
Under review as a conference paper at ICLR 2027

PoisonedEar: Knowledge Poisoning Attacks on Retrieval-Augmented Generation Multimodal Reasoning in Audio-Centric Language Models

Abstract

Recent advances in multimodal retrieval-augmented generation (RAG) allow multimodal large language models (MLLMs) to incorporate external knowledge at inference time. Although audio-centric RAG is increasingly deployed in applications such as voice assistants, meeting summarization, and financial audio analysis, its vulnerability to knowledge poisoning through audio retrieval remains largely unexplored. We present PoisonedEar, a knowledge-poisoning attack that exposes this vulnerability in audio-centric RAG systems. Under a black-box threat model, PoisonedEar induces MLLMs from diverse model families to generate attacker-controlled responses to audio queries by injecting only a small number of malicious audio–text pairs into the knowledge database. Our experiments demonstrate that this threat is both practical and severe. In decision-critical scenarios, a handful of poisoned entries can cause an MLLM to raise false alarms or suppress legitimate warnings, achieving attack success rates of up to 80%. Moreover, representative defenses provide only limited protection against the attack. These findings establish knowledge poisoning as an effective and stealthy threat to audio-centric RAG systems and highlight the need for more robust retrieval and knowledge-base safeguards. The code is available at https://github.com/PoisonedEarProject/PoisonedEar

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Knowledge Poisoning, Audio RAG

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_GzyHOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Author provided Github link (https://github.com/PoisonedEarProject/PoisonedEar) contains commit logs that explicitly show the author initial and last name ("SH Zhang"). I deem this as a failure to abide by the anonymity guideline and hence request the AC to give desk rejection.

Critical strengths

N/A

Critical weaknesses

Author provided Github link (https://github.com/PoisonedEarProject/PoisonedEar) contains commit logs that explicitly show the author initial and last name ("SH Zhang"). I deem this as a failure to abide by the anonymity guideline and hence request the AC to give desk rejection.

Questions for authors

N/A

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

This paper investigates knowledge poisoning in audio-centric retrieval-augmented generation (RAG) systems. It introduces PoisonedEar, an attack that pairs proxy audio recordings with malicious captions to steer a victim model toward an attacker-chosen interpretation. To construct misleading captions that remain consistent with the audio’s acoustic characteristics, it proposes Cross-Domain Acoustic Bridging (CDAB), which attributes the sounds to a different semantic context. The experiment covers four audio datasets and 16 multimodal models across 4 model families, with comparisons of caption construction strategies and analyses of retrieval configurations and poisoning rates. It also investigates how different defense strategies mitigate the attack.

Critical strengths
  1. The paper is easy to follow. It also cleary defines the threat model.

  2. Some of the motivations of this paper is well motivated and validated. Figure 7 compares direct answering with Audio-RAG and shows improved model performance with retrieval, motivating the study of poisoning attacks on Audio-RAG systems.

Critical weaknesses
  1. The three challenges in the audio-centric knowledge poisoning pipeline developed by the authors are overstated, as they are also common in other multimodal pipelines, e.g., image-based pipelines. For C1, image queries can also be unknown in advance and vary in viewpoint, lighting, and background. For C2, a MLLM can understand the input image and reject conflicting retrieved text. For C3, multiple consistent poisoned entries can reinforce a misleading interpretation in other modalities as well.

  2. There are some conflicts between the experimental results and the definitions of the metrics, and some experimental results are suspected to have been fabricated. Here I list some examples:

    • By the definition in Appendix B.3, Recall@k is the ratio between the number of adversarial samples retrieved in the top-k results and the total number of injected adversarial samples. Let ∣Pi∣|\mathcal{P}_i| be the number of injected adversarial samples. Then Recall@k≤min⁡{1,k/∣Pi∣}\mathrm{Recall@k} \leq \min \lbrace 1, k/|\mathcal{P}_i| \rbrace. Why can Recall@5 in Table 2 reach such high values?

    • Also by the definition in Appendix B.3, Precision@k is the ratio between the number of adversarial samples retrieved in the top-k results and kk, averaged over queries. Therefore, Overall ASR≤ASR−R≤min⁡{1,k×P@k}\mathrm{Overall\ ASR} \leq \mathrm{ASR-R} \leq \min \lbrace 1, k \times \mathrm{P@k} \rbrace. Why do the P@5 and Overall ASR values for Clotho and AudioSet in Table 5 conflict with this relation? For example, P@5 = 0.09 and Overall ASR = 0.87 for GPT-4o-audio on Clotho, but 5×0.09=0.45<0.875 \times 0.09 = 0.45 < 0.87.

    • In Figure 12, the Clotho values match 0.7×WavCaps+0.3×AudioSet0.7 \times \mathrm{WavCaps} + 0.3 \times \mathrm{AudioSet}, while the ESC-50 values match 0.2×WavCaps+0.8×AudioSet0.2 \times \mathrm{WavCaps} + 0.8 \times \mathrm{AudioSet}, followed by poisoning-rate-specific scaling and clipping to [0,1][0,1]. This concern is also supported by the code implementation in the provided repository.

Questions for authors
  1. Since the poisoned data is injected into the retrieval corpus, then where do the evaluation queries come from, and how many are used for each dataset? I don't find these details in the paper.

  2. Please address the concern in weakness 2

Reviewer_87zyOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies knowledge poisoning in audio-centric RAG, where a query audio clip retrieves similar audio-text pairs whose captions condition an audio MLLM. It identifies three obstacles for an attacker: future query recordings are unknown, the MLLM can hear the audio and reject captions that contradict it, and a single malicious caption may not be enough to change the model’s answer. PoisonedEar injects several recordings of the source sound with captions produced by Cross-Domain Acoustic Bridging (CDAB). CDAB describes the audio’s low-level acoustic properties and reframes them as evidence for an attacker-chosen target, so the caption stays consistent with what the model hears. The experimental results show that MLLMs can be steered by consistent retrieved context. However, the protocol does not test unseen recordings, a fixed attacker-chosen target or a small injection budget, and it leaves out the most natural audio-aware defence, so how serious the threat is in practice remains unclear.

Critical strengths
  1. The ablation of CDAB is clean and isolates a real audio-specific effect. The naive baseline and PoisonedEar inject identical audio, so retrieval is the same and only the captions differ.
  2. The evaluation is broad, and the framework is released. It covers 16 MLLMs across the GPT, Gemini and open-source families, four retrievers, four datasets, five poisoning rates, a top-k sweep and four defences. The open-source AudioRAG framework is useful for follow-up work on defences.
Critical weaknesses
  • The experimental protocol does not test the threat model the paper claims. Unseen recordings (Challenge 1) are not evaluated. The paper never states that victim queries are disjoint from the poisoned entries or come from different recording sources. If they do not, retrieval of the poisoned audio is close to guaranteed.

  • How attack success is measured is underspecified, and the reported metrics are internally inconsistent. For example, the success criterion is undefined. The paper never says how a response is judged “consistent with R_i”. It also does not give the number of victim queries per class, and it reports no variance.

  • The defence evaluation overstates the attack’s resilience and leaves out the obvious audio-aware defence.

  • Violate the anonymisation policy.

Questions for authors

Are victim queries disjoint from the poisoned entries? Do they come from different recordings or sources than the proxy anchors?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.