A Tandem Framework Balancing Privacy and Security for Voice User Interfaces

Ranya Aloufi, Hamed Haddadi, David BoylearXiv 2021 · arXiv

As voice-activated services like Alexa or banking-by-voice become ubiquitous, they face two conflicting challenges: protecting user privacy and ensuring system security. This paper explores the inherent tension between privacy (hiding a speaker's identity) and security (detecting fraudulent voice clones). The authors propose a "tandem framework" that jointly evaluates anti-spoofing and authentication models, revealing that current voice anonymization techniques are often indistinguishable from malicious spoofing attacks to modern security systems.

The Privacy-Security Duality

The fundamental problem identified is that the fields of voice privacy and security have largely evolved in parallel, despite using the same underlying technology.

  1. Privacy: Users want to interact with services without revealing their biometric identity, gender, or emotional state. This is often achieved through Voice Conversion (VC), which transforms a user's voice into a "pseudo-speaker."
  2. Security: Systems must prevent spoofing attacks, where an attacker uses VC or speech synthesis to impersonate a legitimate user.
The Voice Conversion (VC) process used for anonymization, illustrating the training and testing phases of a conversion model.

A "benign" user trying to protect their privacy uses the same VC technology as a "malicious" attacker trying to bypass security. Consequently, a robust security system might accidentally block a privacy-conscious user, while a privacy tool might inadvertently provide a blueprint for a security breach.

The Tandem Framework

The authors introduce a Tandem Framework to evaluate the interaction between security and privacy. This framework consists of two main components:

  • Verification-to-Disguise (V2D): A spoofing detector (Countermeasure or CM) that acts as a gatekeeper to determine if a voice is genuine human speech or synthetically generated/converted.
  • Verification-to-Anonymization (V2A): An Automatic Speaker Verification (ASV) system that checks if the voice matches a specific target identity.
Spoofing attack scenarios against Automatic Speaker Verification (ASV) systems, distinguishing between non-proactive and adversarial (black/grey/white-box) attacks.

By placing the CM before the ASV in a cascaded pipeline, the authors can measure the tandem detection cost function (t-DCF)—a metric that accounts for the errors made by both systems. They tested five different VC models (ranging from baseline challenges to advanced GAN-based architectures) to see how they performed as both privacy tools and potential spoofing threats.

Experimental Setup and Datasets

The study utilizes four major benchmark datasets to ensure broad applicability:

  • VoxCeleb: Used to train and evaluate the authentication (ASV) systems.
  • VCTK: Used primarily to train the anonymization (VC) systems.
  • VCC2020: A multilingual database used to evaluate voice conversion across different languages.
  • ASVspoof2019: The standard benchmark for logical access attacks, used to train and test spoofing countermeasures.

Key Insights and Results

The "Artifact" Problem

Even advanced voice conversion models leave subtle digital "artifacts" or inconsistencies in the spectral envelope. While these might be imperceptible to humans, modern anti-spoofing systems (CMs) are highly sensitive to them. Consequently, anonymized voices intended for privacy often result in high Equal Error Rates (EER) in the CM, meaning the security system flags them as "not natural human speech" and blocks the user.

Balancing Objectives

If a VUI implements a very "strong" security gatekeeper, it effectively kills the utility of privacy-preserving anonymization. The authors demonstrated that by using the t-DCF metric, they could find an "optimal threshold" for security systems that balances the need to catch attackers while still allowing some level of voice transformation.

Robustness Through Loss Functions

The authors found that using Sigmoid and P2SGrad loss functions in the anti-spoofing models provided better robustness against "unknown" attacks—voices converted by models the system hadn't seen during training.

Future Directions

The paper points toward a need for "configurable privacy" systems where users can choose their level of anonymization based on the specific service they are accessing. Future work should also incorporate adversarial examples (adding imperceptible noise to intentionally fool models) and explore online watermarking as a way to verify voice integrity without storing raw biometric data.

Summary updated

Abstract

Speech synthesis, voice cloning, and voice conversion techniques present severe privacy and security threats to users of voice user interfaces (VUIs). These techniques transform one or more elements of a speech signal, e.g., identity and emotion, while preserving linguistic information. Adversaries may use advanced transformation tools to trigger a spoofing attack using fraudulent biometrics for a legitimate speaker. Conversely, such techniques have been used to generate privacy-transformed speech by suppressing personally identifiable attributes in the voice signals, achieving anonymization. Prior works have studied the security and privacy vectors in parallel, and thus it raises alarm that if a benign user can achieve privacy by a transformation, it also means that a malicious user can break security by bypassing the anti-spoofing mechanism. In this paper, we take a step towards balancing two seemingly conflicting requirements: security and privacy. It remains unclear what the vulnerabilities in one domain imply for the other, and what dynamic interactions exist between them. A better understanding of these aspects is crucial for assessing and mitigating vulnerabilities inherent with VUIs and building effective defenses. In this paper,(i) we investigate the applicability of the current voice anonymization methods by deploying a tandem framework that jointly combines anti-spoofing and authentication models, and evaluate the performance of these methods;(ii) examining analytical and empirical evidence, we reveal a duality between the two mechanisms as they offer different ways to achieve the same objective, and we show that leveraging one vector significantly amplifies the effectiveness of the other;(iii) we demonstrate that to effectively defend from potential attacks against VUIs, it is necessary to investigate the attacks from multiple complementary perspectives(security and privacy).
BibTeX
@misc{aloufi2021tandem,
  author = {Ranya Aloufi and Hamed Haddadi and David Boyle},
  title = {{A Tandem Framework Balancing Privacy and Security for Voice User Interfaces}},
  year = {2021},
  eprint = {2107.10045},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2107.10045},
}
Figure