Scenario of Use Scheme: Threat Model Specification for Speaker Privacy Protection in the Medical Domain

Mehtab Ur Rahman, Martha Larson, Louis ten Bosch, Cristian Tejedor GarcíaarXiv 2024 · DOI · OpenReview

Abstract

Speech recordings are being more frequently used to detect and monitor disease, leading to privacy concerns. Beyond cryptography, protection of speech can be addressed by approaches, such as perturbation, disentanglement, and re-synthesis, that eliminate sensitive information of the speaker, leaving the information necessary for medical analysis purposes. In order for such privacy protective approaches to be developed, clear and systematic specifications of assumptions concerning medical settings and the needs of medical professionals are necessary. In this paper, we propose a Scenario of Use Scheme that incorporates an Attacker Model, which characterizes the adversary against whom the speaker's privacy must be defended, and a Protector Model, which specifies the defense. We discuss the connection of the scheme with previous work on speech privacy. Finally, we present a concrete example of a specified Scenario of Use and a set of experiments about protecting speaker data against gender inference attacks while maintaining utility for Parkinson's detection.

Introduction

Voice recordings are increasingly utilized as "digital biomarkers" to detect and monitor neurodegenerative diseases such as Parkinson’s and Alzheimer’s. However, speech data is inherently sensitive, containing information about a speaker's gender, age, emotional state, and identity. While voice anonymization techniques exist, the research community lacks a standardized language to describe these protections. Often, studies fail to explicitly state the assumptions regarding an attacker's knowledge or the specific medical "utility" that must be preserved. This ambiguity hinders the comparison of privacy solutions and limits trust among medical professionals. To address this, we introduce the Scenario of Use Scheme, a framework designed to bridge the gap between technical privacy research and clinical requirements.

The Scenario of Use Scheme

The core contribution of this work is a standardized framework for describing speaker privacy in medical contexts. The scheme decomposes privacy scenarios into three critical dimensions for both the Attacker and the Protector models:

  1. Objective: Defines what the attacker seeks to obtain and what the protector intends to conceal (Defense Objective) versus retain (Utility Objective).
  2. Opportunity: Specifies the level of access each party has to the target data and the opposing party's systems (e.g., black-box vs. white-box access).
  3. Additional Resources: Accounts for external data, knowledge, and computing power available to carry out the attack or defense.
The Scenario of Use Scheme framework, defining Attacker and Protector Models across three dimensions: Objective, Opportunity, and Additional Resources.

By formalizing these dimensions, researchers can move beyond simple "attacker vs. defender" logic toward a nuanced specification of the clinical pipeline. This standardization also allows for the re-expression of existing work, such as the VoicePrivacy Challenge, making fundamental differences in assumptions more transparent.

Experimental Methodology

To demonstrate the scheme's utility, we conducted a proof-of-concept experiment focusing on protecting speaker gender while maintaining the ability to detect Parkinson’s disease (PD). We utilized the NeuroVoz corpus, which contains Castillian Spanish speech from 108 participants, including healthy controls (HC) and individuals with PD. The dataset was split into a training set of 1,141 recordings and a test set of 210 recordings, using Mel-scale spectrograms as primary features.

The experiment employed Projected Gradient Descent (PGD) to treat privacy as an adversarial machine learning task. The protector applies calculated perturbations to the audio data to "fool" a gender classifier (the attacker) while attempting to preserve the accuracy of a diagnosis classifier (the utility).

Results and Discussion

The experimental results highlight the inherent trade-off between privacy and utility in medical speech processing. The protection mechanism was highly effective at concealing gender; the gender classifier’s performance dropped from an AUC of 0.88 to 0.51, effectively reaching the level of random guessing (0.50).

Experimental results comparing original and perturbed data across various performance metrics for Parkinson's diagnosis.

However, this privacy gain came at a cost to diagnostic utility. As shown in the table above, the Parkinson’s diagnosis accuracy decreased from 87.66% to 80.97%. While a 6.69% drop is measurable, it represents a significant step toward achieving near-total gender privacy in a clinical setting. Future work should focus on refining perturbation techniques to minimize this utility gap.

Conclusion

The Scenario of Use Scheme serves as a vital tool for "data bookkeeping," enabling clearer communication between computer scientists, data managers, and medical professionals. By explicitly defining threat models and utility requirements, the scheme facilitates the development of privacy-preserving technologies that are both robust against adversaries and reliable for clinical diagnosis.

Summary updated

Abstract

Speech recordings are being more frequently used to detect and monitor disease, leading to privacy concerns. Beyond cryptography, protection of speech can be addressed by approaches, such as perturbation, disentanglement, and re-synthesis, that eliminate sensitive information of the speaker, leaving the information necessary for medical analysis purposes. In order for such privacy protective approaches to be developed, clear and systematic specifications of assumptions concerning medical settings and the needs of medical professionals are necessary. In this paper, we propose a Scenario of Use Scheme that incorporates an Attacker Model, which characterizes the adversary against whom the speaker's privacy must be defended, and a Protector Model, which specifies the defense. We discuss the connection of the scheme with previous work on speech privacy. Finally, we present a concrete example of a specified Scenario of Use and a set of experiments about protecting speaker data against gender inference attacks while maintaining utility for Parkinson's detection.
BibTeX
@misc{rahman2024scenario,
  author = {Mehtab Ur Rahman and Martha Larson and Louis ten Bosch and Cristian Tejedor García},
  title = {{Scenario of Use Scheme: Threat Model Specification for Speaker Privacy Protection in the Medical Domain}},
  year = {2024},
  eprint = {2409.16106},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2409.16106},
}
Figure