Intuitive Privacy from Acoustic Reach: A Case for Networked Voice User-Interfaces

Tom Bäckström, Sneha Das, Pablo Pérez Zarazaga, Johannes Fischer, Rainhard Dieter Findling, Stephan Sigg, Le Ngu Nguyen2021 · DOI

TL;DR

As voice interfaces become more integrated into our lives, especially in multi-device systems, privacy concerns are amplified. This paper proposes using the "acoustic reach" of speech – how far a sound travels – as an intuitive mechanism to define and manage privacy in networked voice user interfaces, requiring only modest technological extensions.

What Problem Does This Paper Address?

The proliferation of voice-enabled devices, from smart speakers to in-car assistants, has outpaced our understanding and mitigation of their privacy implications. While current systems often rely on single devices or cloud-based processing, the trend is towards collaborative, multi-device networks that offer a more seamless user experience. However, this increased connectivity raises significant privacy questions: which devices should have access to speech data, and how can we ensure that this access is authorized and respects user expectations? The paper argues that existing approaches to privacy in these systems are insufficient and that a more intuitive, user-centric approach is needed.

Methods

The core idea proposed is to leverage the acoustic reach of speech as a natural boundary for privacy. This means that a device should only be allowed to process or transmit speech if it is within the audible range of the speaker, mimicking how humans intuitively manage privacy in physical spaces.

The authors suggest that this can be implemented through:

  • Acoustic Fingerprinting: Using techniques similar to those in [7-10] to derive unique "fingerprints" from audio signals. These fingerprints can be compared between devices to determine if they are within the same acoustic space without revealing the actual content of the speech.
  • Implicit Access Rules: Designing systems where collaboration between devices is automatically enabled if they are within the same acoustic reach. This contrasts with explicit opt-in mechanisms for every interaction.
  • Threat Analysis: Examining potential attack vectors, such as injecting a single malicious device into a network, and how the acoustic reach concept can mitigate these threats.

Datasets

Not discussed in the paper. The paper focuses on conceptual exploration and experimental validation of the acoustic reach principle rather than training on large datasets.

Key Insights and Ideas

  • Intuitive Privacy: Users naturally adjust their speaking volume and behavior to manage privacy in different social contexts (e.g., speaking softly in a living room, whispering secrets in a cafeteria). The acoustic reach concept aligns with this innate understanding.
  • Acoustic Reach as a Feature: Instead of treating privacy as an afterthought or a complex configuration setting, it can be an inherent feature of the voice interface, defined by the physical properties of sound propagation.
  • Modest Technological Requirements: Implementing acoustic reach-based privacy doesn't necessitate entirely new technologies. Existing methods for audio fingerprinting and proximity detection can be adapted.
  • Ethical Considerations in Design: The paper emphasizes the importance of ethical considerations in engineering, particularly when dealing with sensitive data like speech. It argues against purely objective measurements for ethical boundaries, favoring scenario-based analysis.

Key Results

The paper presents two experimental evaluations:

  1. Distance and Fingerprint Similarity: Experiments showed that the similarity of audio fingerprints between devices decreases significantly as the physical distance between them increases. This supports the feasibility of using audio fingerprinting to determine acoustic proximity.
  2. Secure Device Pairing: In a simulated device pairing scenario, the audio fingerprint similarity between devices within the same room was considerably higher than that between a device in the room and an "attacker" device outside the room (even with the door open). This suggests acoustic reach can help distinguish legitimate from malicious devices during pairing.

Limitations and Future Work

  • Ethical Measurement Challenges: The authors acknowledge that directly measuring ethical issues is problematic and that their approach relies on scenario-based discussions.
  • Usability Studies: Future work might involve designing usability studies that do not pose ethically unsustainable risks, particularly when evaluating the performance of privacy-enhancing systems.
  • Broader Application: The paper suggests that this approach could lead to more intuitive and privacy-friendly voice user interfaces, potentially reducing vendor lock-in and fostering competition by enabling interoperability between devices from different manufacturers.
  • Decentralized Logging: For auditing purposes, the paper mentions the potential for decentralized consensus mechanisms (like blockchain) or server-based solutions for logging access rights.

Summary updated

Abstract

The effect that advances in voice interface technologies have on privacy has not yet received the attention it deserves. Systems in which multiple devices collaborate to provide a unified user-interface amplify those worries about privacy. We discuss ethical implications of voice enabled devices on privacy in typical scenarios at home, office, in a car and in the public. From our findings, it follows that the reach of voice can be exploited as a feature to intuitively define the extent of privacy. In particular, the acoustic reach of speech signals can serve as a feature for designing privacy-gentle voice user-interfaces which are intuitive to use. We argue that this approach poses reasonable technological requirements and establishes a natural experience of privacy which confirms intuitive perception.
BibTeX
@misc{backstrom2021intuitive,
  author = {Tom Bäckström and Sneha Das and Pablo Pérez Zarazaga and Johannes Fischer and Rainhard Dieter Findling and Stephan Sigg and Le Ngu Nguyen},
  title = {{Intuitive Privacy from Acoustic Reach: A Case for Networked Voice User-Interfaces}},
  year = {2021},
  doi = {10.21437/spsc.2021-12},
}
Figure