Privacy-preserving voice anti-spoofing using secure multi-party computation
(3 minutes introduction)
|Oubaïda Chouchane (EURECOM, France), Baptiste Brossier (EURECOM, France), Jorge Esteban Gamboa Gamboa (EURECOM, France), Thomas Lardy (EURECOM, France), Hemlata Tak (EURECOM, France), Orhan Ermis (EURECOM, France), Madhu R. Kamble (EURECOM, France), Jose Patino (EURECOM, France), Nicholas Evans (EURECOM, France), Melek Önen (EURECOM, France), Massimiliano Todisco (EURECOM, France)|
In recent years the automatic speaker verification (ASV) community has grappled with vulnerabilities to spoofing attacks whereby fraudsters masquerade as enrolled subjects to provoke illegitimate accepts. Countermeasures have hence been developed to protect ASV systems from such attacks. Given that recordings of speech contain potentially sensitive information, any system operating upon them, including spoofing countermeasures, must have provisions for privacy preservation. While privacy enhancing technologies such as Homomorphic Encryption or Secure Multi-Party Computation (MPC) are effective in preserving privacy, these tend to impact upon computational capacity and computational precision, while no available spoofing countermeasures preserve privacy. This paper reports the first solution based upon the combination of shallow neural networks with secure MPC. Experiments performed using the ASVspoof 2019 logical access database show that the proposed solution is not only computationally efficient, but that it also improves upon the performance of the ASVspoof baseline countermeasure, all while preserving privacy.