acceptodds
Under review as a conference paper at ICLR 2027

What Spatial Cues Do Multichannel Replay Detectors Rely On? A Physically Interpretable Analysis via Relative Transfer Functions

Abstract

Replay attacks pose a major security threat to widely deployed voice interaction systems. Multichannel detectors typically exploit spatial information between microphones, but their reliance on specific spatial features remains unclear. We use physically interpretable relative transfer functions (RTFs) as control variables to adjust inter-channel magnitude ratios and phase differences. Without accessing model internals, we select transferable RTF templates using only detection scores from a small set of recordings. These templates can induce replay-to-genuine misclassification on recordings and some models not used in selection, while transcripts and speaker representations remain highly consistent after processing. These findings directly connect model decisions to interpretable physical features across channels, providing insight into multichannel replay detectors and a new perspective on their robustness.

Primary area:
interpretability and explainable AI
Keywords:
multichannel replay detection, relative transfer functions, model interpretability, adversarial robustness, black-box attacks

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_GYGAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper made an experimental study on the spatial cues that multichannel reply detectors reliy on.

Critical strengths

An experimental framework for the interpretability study of multichannel replay detectors is designed.

Critical weaknesses
  1. The novelty of the study is not significant. The proposed experimental framework is based on existing techniques. The evaluated detectors are also existing ones.
  2. The experiment section gave plenty of experimental results. However, as an experimental study on interpretability, more insightful analysis is expected and there are no valuable conclusions which can help to overcome the main challenges of replay detection.
Questions for authors

Some introductions on existing techniques can be abbreviated. For example, it is difficult to discriminate which parts are exsiting methods are which parts are the contributions of the paper in Section 3.2.

Reviewer_cgL5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposed a detection strategy against replay attacks via physically interpretable relative transfer templates (RFTs). The use of RTF is mainly for adjusting the magnitude ratio and phase difference to analyze the detector dependence on inter-channel structure. Experiments have been conducted on the ReMASC dataset and different detectors.

Critical strengths
  • The paper explores an important research problem on replay attack detection.
Critical weaknesses
  • Honestly, the paper is very poorly written and hard to follow. After reading it multiple times, I still can not understand its main novelty and contributions. The writing should be improved and the flow needs to be more clear. The paper should speak for itself and make it easy for the reader to understand its novelty and science more clearly.
  • The paper’s novelty is the main question here. There’s a related work section but it doesn’t discuss any previous baselines that the paper could compare against. So a clarification of how the proposed approach is advancing the research landscape of replay attacks is very much needed in this paper.
  • The evaluation is also hard to follow, what are D1-D4? What are the J, A, P replay? A clear description of the experimental setup is also needed to understand on which attacks the methods have been evaluated, metrics, baselines, and more datasets to showcase to what extent this method is scalable.
  • Overall, the paper really needs much work to make it a strong research contribution. Clear writing, research motivation, experimental setup, evaluation, and insights are required to make this a strong research paper.
Questions for authors
  • What's the scientific novelty of this paper?
  • What are the existing baselines? How the proposed approach is advancing the research field on replay attacks?
  • Please explain and provide more details on the experimental setups?
  • Please clarify the research findings and insights
Reviewer_HSq2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper uses relative transfer functions to probe multichannel replay detectors through joint, magnitude-only, and phase-only channel reconstruction. Empirical templates and score-selected parameterized templates change decisions across seven detectors and four arrays, with some cross-model transfer. Separate full-utterance experiments measure transcript and speaker-embedding consistency.

Critical strengths

The RTF parameterization makes the interventions easier to interpret than an unrestricted waveform perturbation. Evaluation across seven detectors and four arrays reveals useful differences in sensitivity, and reusing fixed templates on held-out recordings and other models strengthens the robustness findings.

Critical weaknesses
  1. The attack is evaluated by digitally modifying existing ReMASC recordings. It is unclear how a playback attacker could impose the required per-microphone RTFs. There is no playback-and-rerecording or held-out environment/device evaluation, limiting the practical significance of the high conversion rates.

  2. Attack-aware defenses are not evaluated; RTF augmentation is only mentioned as future work. A simple retraining baseline, tested against newly searched templates at matched clean false-positive rates, would clarify whether the vulnerability persists once the defender knows the attack.

Questions for authors
  1. How can a playback attacker realize the required RTFs, and does the attack survive playback and rerecording?

  2. Does RTF-augmented training remain vulnerable after attack templates are re-selected against the defended detector?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.