|
|
Testing the Robustness of a Voice Biometrics System against Deepfakes
Reš, Jakub ; Firc, Anton (oponent) ; Malinka, Kamil (vedoucí práce)
Topic of this paper is a methodology of testing the robustness of a voice biometric system against deepfakes. The main problem currently lies in insufficient coverage of testing against the presentation attack using deepfakes in ISO/IEC standards. The aim of this thesis is to cover the hole, resulting from emergence of deepfake technology, by proposing an extended methodology, based on the existing one, that focuses on fixing the issue. The solution of proposed problem started by studying the state of the art for deepfakes and standard practices of biometric system testing. Second, I proposed and documented a method of testing the voice biometric system. The test was designed as a scenario, where the Phonexia voice biometric system is used as a remote verification tool for the voice-as-a-password use-case. For the purpose of demonstration, the online publicly available dataset was used. On top of test design, I set a non-standard metric for the test evaluation to show possibilities of focus on different kinds of deepfakes. After carrying out tests and evaluating results, I formulated the procedure into a generic repeatable methodology, containing practices and recommendations. The contribution of this work lies in incorporating deepfakes into the existing standard methodologies of testing a biometric systems, hence forming and demonstrating a repeatable methodology.
|
host ::
RSS.