In this paper, we propose a framework for evaluating the robustness of speech pseudonymization methods. Among privacy-protecting methods, signal processing-based and machine learning-based methods have been proposed as pseudonymization methods. Although most studies evaluate the pseudonymization performance of a method, the consideration of irreversibility, which is as important as performance, has been absent. While there are studies taking account of threats from the disclosure of partial information of a pseudonymization scheme, only few have discussed irreversibility for the case of malicious attacks, namely decryption attacks. Thus, we demonstrate irreversibility by evaluating the robustness of pseudonymization methods against decryption attacks in this paper. A decryption attack scenario is assumed that is advantageous to attackers because it allows access to the internal design and the pseudonymized speech generated from a pseudonymization system. From there, attackers try to build a system to decrypt pseudonymized speech to reveal the identity behind the original speech. In our experiments, we evaluate our previously proposed pseudonymization methods using single or multiple signal processing-based speech modification methods. The results demonstrate that “single” ones are vulnerable to decryption attacks, whereas “multiple” ones greatly improve the robustness against such attacks.
Cite as: Kai, H., Takamichi, S., Shiota, S., Kiya, H. (2022) Robustness of Signal Processing-Based Pseudonymization Method Against Decryption Attack. Proc. The Speaker and Language Recognition Workshop (Odyssey 2022), 287-293, doi: 10.21437/Odyssey.2022-40
@inproceedings{kai22_odyssey,
author={Hiroto Kai and Shinnosuke Takamichi and Sayaka Shiota and Hitoshi Kiya},
title={{Robustness of Signal Processing-Based Pseudonymization Method Against Decryption Attack}},
year=2022,
booktitle={Proc. The Speaker and Language Recognition Workshop (Odyssey 2022)},
pages={287--293},
doi={10.21437/Odyssey.2022-40}
}