acceptodds
Under review as a conference paper at ICLR 2027

Same Audio, Different Boundaries: Inferring Context, Drawing the Line, Guarding Privacy

Abstract

udio-capable models are increasingly used in applications such as personal assistants, raising privacy concerns that extend beyond conventional personally identifiable information (PII) to whether information is disclosed appropriately. This contextual dimension remains underexplored in audio scenarios. We define Contextual Audio Privacy (CAP) from an information flow perspective, where permissible disclosure is determined by the interaction context. To study CAP systematically, we construct CAP-data, a dataset of 9,000 instances spanning three settings with progressively richer contextual constraints. Empirical results show that privacy violations are widespread across the 13 evaluated audio-capable models. Privacy behavior varies substantially with requester authorization and intended use, while downstream task has a much smaller effect. Motivated by these findings, we propose FlowInspect, a runtime enforcement controller that separates privacy decisions from task execution. FlowInspect parses the relevant interaction context into audio speaker, requester role, and intended use, and maps it through a deterministic policy engine to an explicit disclosure level. Experiments show that FlowInspect reduces average privacy violation rate (PVR) from 56.82% to 31.27% across four target models while retaining 78.18% of Vanilla task successes. A 0.6B context parser retains 93.7% of the 8B parser's exact match accuracy with substantially lower memory cost.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Audio Privacy, Contextual Privacy, Audio-Capable Models, Information Flow, Privacy Enforcement, AI Safety

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

The paper introduces a 9,000-instance benchmark for contextual audio privacy and FlowInspect, an external controller that reduces privacy violations across four audio models while retaining much of their task utility.

Critical strengths

The paper is easy to follow, and the controller can be applied to different audio models without retraining them.

Critical weaknesses
  1. Much of the motivation overlaps with VoxPrivacy, which already studies how disclosure should depend on the speaker and interaction context. The new scenarios and partial-disclosure setting are useful extensions, but their added value needs to be clearer. FlowInspect also infers authorization from the request text without verifying it, so the benign-request assumption leaves an important limitation on its practical use.
  2. The PC-TSR results appear inconsistent with the definition in Appendix A.5. For Gemini-3.7-Flash on Speech, LIMITED/NONE cases make up 75% of the data, so a PVR of 69.87% leaves at most about 6.84% of those cases both compliant and potentially successful. Table 1 reports 92.44%.
Questions for authors

See Weakness.

Reviewer_i186OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Contextual Audio Privacy (CAP), a contextual-integrity style formulation in which the permitted disclosure level (NONE / LIMITED / FULL) of private audio content depends on the audio speaker, requester role, and intended use. It builds CAP-data: 9,000 audio-instruction pairs over three scenarios (Speech, Bystander, Conference), obtained by pairing 400 audio items with LLM-generated instructions that enumerate requester x use x task combinations; the permitted level is assigned by an author-defined policy table. Thirteen audio-capable models are evaluated with an LLM judge (Qwen3.6-27B); average Privacy Violation Rate (PVR) is 54% under a vanilla prompt. The paper then proposes FlowInspect: a fine-tuned Qwen3-0.6B parser that extracts (speaker, requester, use) from the instruction text, a deterministic policy lookup, and a disclosure instruction prepended to the target model's prompt. FlowInspect lowers average PVR across four models from 56.8% to 31.3% while retaining 78% of vanilla task successes.

Critical strengths

-Clear problem statement and a clean separation of policy decision from task execution; the three scenarios (especially Conference with stage-based access) are a sensible instantiation of contextual integrity for audio.

-The evaluation is broad (13 models, 4 targets for enforcement, parser scale study) and the prompts and policies are fully listed in the appendix.

Critical weaknesses

-A threat model may make the problem close to trivial. Context (who the requester is, what the purpose is) is read from the requester's own instruction text; the paper explicitly assumes "benign requests without adversarial manipulation". However, a policy engine that trusts a self-declared "I am the treating doctor" provides no privacy protection against anyone willing to type a different role.

-Averaged over the four target models, the simplest baseline (Policy Prompt) reaches 30.6%=(15.61%+28.56%+36.7%+41.38%)/4 PVR, and the Context Parser variant 27.9%, both lower than FlowInspect's 31.3%. FlowInspect is the worst of the three enforcement variants on the privacy metric. The abstract and Section 6.2 compare FlowInspect to Vanilla only and never state this.

Questions for authors

Q1. Why does FlowInspect have higher PVR than Context Parser on Gemini-3.7-Flash (25.4 vs 21.1) and Doubao (34.0 vs 20.7), given that FlowInspect removes one inference step from the target model? Does the "trusted disclosure" prefix make models more willing to disclose when the level is FULL or LIMITED?

Q2. Oracle Boundary improves PVR over FlowInspect by only 0.1-5 points, even though the parser's exact match is 72.6% (and about 60% on Bystander). Does this mean parser errors mostly land on label pairs that map to the same disclosure level?

Q3. Which LLM generated the instructions and which TTS voices were used for Conference? Are the 1,800 parser training instructions generated by the same LLM and template as the test instructions? If so, parser accuracy on held-out instructions is an in-distribution number.

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

The paper concerns privacy preservation in the context of spoken input to audio capable models. A 'contextual audio privacy' (CAP) formulation is proposed whereby speaker, requester, content, use and downstream tasks are provided as relevant factors. Three settings are considered (speaker, bystander and conference), for which CAP-data is constructed for five common audio tasks to examine how contextual factors affect privacy behavior across settings and tasks. FlowInspect is proposed as a runtime enforcement controller to separate privacy decisions from task execution. Privacy violations are suggested to occur widely across heterogeneous models, and that FlowInspect can reduce the instances of violations for the studied settings.

Critical strengths
  • Privacy preservation considering audio-capable models is a very interesting, important and timely challenge.
Critical weaknesses
  • No threat model is provided. The scenarios as presented are not convincing or sufficient in lieu thereof.
  • The establishment of realistic context and settings is limited in the paper, and very difficult to ascertain from the main text. In particular, it is unclear what the bystander and conference inputs actually look like (some examples would be helpful), or how meaningfully the disclosure boundary can be implemented in terms of identifying who 'requester' would correspond to (connecting back to the previous comment).
  • There is very little by way of insight into how or why the proposed solution works. On the one hand, it seems as if the outcome is very much conditioned and to be broadly expected based on the inputs provided, but without any real effort to establish why the various models' behave or respond as they do.
  • Limitations not properly addressed.
Questions for authors

Can you provide a realistic threat model? Please explain the role of the audio capable model in the contexts provided, and be clear about the requester in particular.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.