acceptodds
Withdrawn from ICLR 2027

Room for Error: Large-Scale Simulation of Over-the-Air Acoustic Attacks

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Studies of the risks facing AI models follow a well trodden path. However, acoustic models provide a unique challenge: utterances exist in both the physical and digital world. However, across the adversarial literature this challenge is either ignored with digital only evaluations, or tested within a single lab environment. This is a methodological and metrological shortcoming that undermines our understanding of risk. Through real-world testing, conceptual discussions, and a novel, high-throughput acoustic simulation framework, we demonstrate a new path for adversarial acoustics. This includes formalizing and operationalizing a novel Dual-Form Signal to Noise Ratio to decouple source stealth from victim attack efficacy, resolving a crucial limitation in current works. We support this with experiments that demonstrate that attackers with coarse room knowledge achieve relative Word Error Rate increases of between 26 and 71%, rising to 94.5% under an oracle upper bound. This work lays the groundwork for repeatable, verifiable research that embraces, rather than abstracts, the acoustic environment.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
adversarial attacks, audio, ASR, automatic speech recognition

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_jocKOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper points out three main problems from existing literature in acoustic attacks and proposes a novel method of evaluating those attacks in a more realistic scenario. The paper analyzes that first, the attack success rate of the digital attack does not guarantee and reflect success rates of real-world OTA attack since there are more noises such as reverberation, attenuation, and spatial geometry. Second, the authors claim that the success rate of OTA depends on the attacker's knowledge of the room. To do so, the authors introduce knowledge gradient, ranging from native to oracle. The results have shown that knowing the exact RIR does not guarantee higher attack success, where Whisper for example Naive FGSM was more effective than RIR-aware FGSM. Third, the authors claim that stealth and effectiveness of the attack should be measured in different parts, meaning SNR for source and victim should be measured separately and calculate the projection cost.

Critical strengths
  1. The paper clearly points out evaluation flaws of the existing literature of OTA acoustic attacks.
  2. Measuring the effectiveness of existing OTA attacks in different RIRs and simulated room environments demonstrate the robustness of the attacks
  3. Knowledge gradient further demonstrates the knowledge of the attackers and measure the effectiveness of the crafted adversarial audio.
Critical weaknesses
  1. While the authors critically points out the main concerns of the evaluation methods of existing attack literature, the paper itself does not have a simulation-to-real-world validation. Previous attack papers either processed the attack in digital-only or small sample of physical environment, the paper itself also evaluate in a simulated acoustic environment, which contradicts the claim. Did the authors consider measuring how close their simulated environment is to the real-world acoustic settings?

  2. Though the introduction of knowledge gradient is interesting, the reviewer believes it did not clearly demonstrate the effectiveness of knowledge gradient. While the attack success rate does not have to be distinctively separated between possessing the knowledge gradient or not, at least the results should have demonstrated some variance through the introduction of knowledge gradient. One interesting result was the Naive FGSM was more effective than RIR-aware FGSM, but did the authors try to interpret why this was the case? Or was RIR-aware truly "RIR-aware"?

  3. The reviewer is concerned about the novelty of the paper. The papers main contribution comes from differential RIR simulation, knowledge gradient, and Dual-Form SNR, while Qin et al. (2019) [1] already evaluated their methods through image source method, different room configuration both white-box and black-box, which is quite similar to what the authors were proposing as their contribution and novelty. If I may have missed, what do the authors believe their main contribution that separates from Qin et al. (2019) [1]?


References

[1] Qin, Yao, et al. "Imperceptible, robust, and targeted adversarial examples for automatic speech recognition." International conference on machine learning. PMLR, 2019.

Questions for authors

Please follow questions from Critical Weaknesses.

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

The paper argues that adversarial attacks on ASR should be evaluated across many simulated rooms rather than purely digitally or in a single physical room. It introduces three things: a four-level “Knowledge Gradient” of attacker room knowledge, a “Dual-Form SNR” that separates stealth at the source from interference at the victim, and a room-simulation framework built on PyRoomAcoustics. Validation is about 8M simulated FGSM/PGD evaluations on Whisper and wav2vec2 across 50 rooms, plus a small white-noise experiment in one physical room.

Critical strengths
  1. The problem definition for over-the-air attacks is interesting and useful.
  2. The evaluation is comprehensive and spans many rooms.
Critical weaknesses
  1. Limited realism. The simulator is not validated against physical measurements. The physical experiment uses only white noise, 10 utterances, and one room, and its layouts are never simulated for comparison.
  2. Section 2 argues that a real over-the-air attacker only hears the speech through the room. Yet all attacks compute δ from the clean source audio x with perfect synchronization.
  3. Limited novelty. The “Knowledge Gradient” is four discrete levels, not a gradient. Blind, Approximate and Oracle FGSM give nearly identical WER, and Blind ≈ Approximate for PGD (Table 4)
Questions for authors

See weaknesses above

Reviewer_nrc6OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a simulation framework to evaluate over-the-air adversarial attacks against automatic speech recognition systems. Its main contributions are (1) a dual-location SNR formulation that separates interference as the speaker and victim locations, (2) a knoweldge gradient describing different levels of environmental information, and (3) an evaluation pipeline built on the room impulse response simulation.

Critical strengths
  • The physical simulations seem promising for evaluating and validating over-the-air acoustic attacks.

  • Broad experimental coverage. The evaluation spans multiple models, attack methods, perturbation budgets, and room configurations.

Critical weaknesses
  • The contribution appears limited and incremental.

The three main ideas build on established work in attacker knowledge, spatial signal measurements, and room simulation. Sampling room conditions also resembles existing RIR-based augmentation and expectation-over-transformation methods. Combining these ideas may provide a useful evaluation framework, but the paper does not clearly show what new insights this combination offers.

  • The positioning against prior work needs clarification and correction.

Section 1 states that prior work used simulated reverberation for attack generation rather than evaluation across a distribution of environments, and Table 3 marks Qin et al. (2019) as not evaluating unknown room geometries. However, that work explicitly uses 1,000 randomly sampled training rooms and a separate set of 100 randomly sampled test rooms, providing a direct precedent for distributional evaluation. Two additional studies are closely related: Imperio (Schönherr et al., ACSAC 2020) optimizes through sampled RIRs without exact target-room knowledge and evaluates physical attacks across three rooms, including configurations without direct line of sight; Metamorph (Chen et al., NDSS 2020) addresses device and channel frequency selectivity through prior channel measurements and domain adaptation. The relationship to Abdullah et al.’s SoK (IEEE S&P 2021) also warrants clearer discussion.

Collectively, these studies challenge the broad characterization that previous research either ignores acoustics or relies on precisely known, single-environment conditions. The authors should distinguish established conceptual insights and evaluation practices from the new formalization, implementation capabilities, and empirical findings contributed here. This would enable a more precise assessment of the work’s value compared to prior research.

  • The paper is not easy to read. The motivation in the abstract and introduction seems too broad compared to the actual problem this paper solves. Some of terms need to be defined or introduced, like "acoustic tax", “information cost”, etc.

  • The physical experiments motivate room-aware evaluation but do not validate the main simulated attack findings.

The physical study uses white noise rather than optimized FGSM or PGD perturbations. It therefore does not establish whether simulated attack rankings or predicted source–victim SNR differences hold under physical playback. A small, representative experiment using optimized perturbations would provide more direct validation.

The physical evaluation also uses only 10 randomly sampled utterances from LibriSpeech clean-100. Random sampling alone does not establish representativeness or adequately control speaker-dependent variation. Reporting the number of speakers and testing a broader range of utterances would strengthen the evidence. An additional corpus could help assess generalization, but would not replace validation of the simulated adversarial results.

Questions for authors
  • How does the proposed framework differ from Qin et al. (2019), which evaluates attacks on 100 independently sampled test rooms? Could you discuss that more?

  • Could you discuss more about the limitations of existing over-the-air acoustic attacks, like Imperio (Schönherr et al., ACSAC 2020) and Metamorph (Chen et al., NDSS 2020) to demonstrate your motivation and contribution compared to them?

  • Which main conclusion from this paper would be difficult to obtain using existing RIR-based simulation and expectation-over-transformation methods? Could you highlight them?

  • The physical experiments use white noise rather than optimized adversarial perturbations. Can you provide a small representative evaluation showing whether the simulated attack rankings and source–victim SNR differences hold under physical playback? What explains the high no-attack WER in the physical experiments?

  • How many speakers are represented in the 10 utterances used for physical evaluation? How sensitive are the results to the selected utterances and speakers?

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

Related papers

Discussion (0)

Sign in to comment.