Privacy-preserving Voice Analysis via Disentangled Representations

Abstract

Voice User Interfaces (VUIs) are increasingly popular and built into smartphones, home assistants, and Internet of Things (IoT) devices. Despite offering an always-on convenient user experience, VUIs raise new security and privacy concerns for their users. In this paper, we focus on attribute inference attacks in the speech domain, demonstrating the potential for an attacker to accurately infer a target user's sensitive and private attributes (e.g. their emotion, sex, or health status) from deep acoustic models. To defend against this class of attacks, we design, implement, and evaluate a user-configurable, privacy-aware framework for optimizing speechrelated data sharing mechanisms. Our objective is to enable primary tasks such as speech recognition and user identification, while removing sensitive attributes in the raw speech data before sharing it with a cloud service provider. We leverage disentangled representation learning to explicitly learn independent factors in the raw data. Based on a user's preferences, a supervision signal informs the filtering out of invariant factors while retaining the factors reflected in the selected preference. Our experimental evaluation over five datasets shows that the proposed framework can effectively defend against attribute inference attacks by reducing their success rates to approximately that of guessing at random, while maintaining accuracy in excess of 99% for the tasks of interest. We conclude that negotiable privacy settings enabled by disentangled representations can bring new opportunities for privacy-preserving applications.

ObscureNet

“…While pervasive sensing and inconspicuous data collection bring about new services, they can lead to major privacy concerns. As shown in previous work, voice data collected by the built-in microphone of virtual personal assistants and mobile phones, often without user's knowledge and consent, can be analyzed to infer their emotional state and mental health condition [3,12]. Similarly, cameras embedded in mobile devices can collect sensitive and personal data, especially in indoor environments [37].…”

“…It can be readily seen that using attribute-specific VAEs improves the F1-score of activity inference from 65.51% obtained by a General VAE to 72.45%. This can be attributed to the fact that having a specific VAE for each public attribute can partly address the imbalance problem in the training data 3 . Unfortunately this comes at the price of increasing the F1-score of gender inference.…”

“…On the other hand, the collected data may contain specific patterns that disclose private attributes associated with people present in the monitored environment, such as their age, gender, race, weight, height, personality traits, and moods. These private attributes can be easily inferred from data using machine learning techniques [3,24,36], leading to serious privacy concerns.…”

Configurable Privacy-Preserving Automatic Speech Recognition

“…Privacy enhancing technology for voice assistive systems is an emerging and important area of research. Aloufi et al describe scenarios whereby attackers can infer a significant amount of private information by observing the output of state-of-art deep acoustic models for speech processing tasks [5]. Nautsch et al,in [6], demonstrate the importance of proposing and developing new privacy-preserving technologies for protecting speakers and speech characterization in voice signals.…”

MegaMind

“…There is a line of work on enhancing the privacy of voice-based systems by eliminating personal features from audio recordings via local preprocessing [44,45]. MegaMind's local speech-to-text conversion also eliminates all voice-based features.…”