Fine-grained, multimodal speech anonymization

31/07/2022 – 01/08/2022 all-day

Offre en lien avec l’Action/le Réseau : – — –/– — –

Laboratoire/Entreprise : Inria Nancy & Lille
Durée : 36 mois
Contact :
Date limite de publication : 2022-07-31

Contexte :
This PhD is part of the “Personal data protection” project of PEPR Cybersécurité, which aims to advance privacy preservation technology for various application sectors. It will be co-supervised by Emmanuel Vincent and Marc Tommasi. The PhD student will have the opportunity to spend time in both the Multispeech and Magnet teams, to collaborate with 9 other research teams in France and with the French data protection authority CNIL, and to contribute to the project’s overall goals including the organization of an anonymization challenge.

Sujet :
Large-scale collection, storage, and processing of speech data poses severe privacy threats [1]. Indeed, speech encapsulates a wealth of personal data (e.g., age and gender, ethnic origin, personality traits, health and socio-economic status, etc.) which can be linked to the speaker’s identity via metadata or via automatic speaker recognition. Speech data may also be used for voice spoofing using voice cloning software. With firm backing by privacy legislations such as the European general data protection regulation (GDPR), several initiatives are emerging to develop and evaluate privacy preservation solutions for speech technology. These include voice anonymization methods [2] which aim to conceal the speaker’s voice identity without degrading the utility for downstream tasks, and speaker re-identification attacks [3] which aim to assess the resulting privacy guarantees, e.g., in the scope of the VoicePrivacy challenge series [4].

The first objective of this PhD is to improve the privacy-utility tradeoff by better disentangling speaker identity from other attributes, and better decorrelating the underlying dimensions. Solutions may rely on suitable generative or self-supervised models [5, 6] or on adversarial learning [7]. The resulting privacy guarantees will be evaluated via stronger attackers, e.g., taking metadata into account.

The second objective is to extend the proposed audio-only approach to multimodal speech (audio, facial video, and gestures). Solutions will exploit existing facial anonymization technology [8]. A key difficulty will be to preserve the correlations between modalities, which are essential for training multimodal voice processing systems.

Depending on the PhD student’s skills, additional directions may also be explored, e.g., evaluating the proposed anonymization solutions in the context of federated learning.

[1] A. Nautsch, A. Jimenez, A. Treiber, J. Kolberg, C. Jasserand, E. Kindt, H. Delgado, M. Todisco, M. A. Hmani, M. A. Mtibaa, A. Abdelraheem, A. Abad, F. Teixeira, M. Gomez-Barrero, D. Petrovska, N. Chollet, G. Evans, T. Schneider, J.-F. Bonastre, B. Raj, I. Trancoso, and C. Busch, “Preserving privacy in speaker and speech characterisation,” Computer Speech and Language, vol. 58, pp. 441–480, 2019.

[2] B. M. L. Srivastava, M. Maouche, M. Sahidullah, E. Vincent, A. Bellet, M. Tommasi, N. Tomashenko, X. Wang, and J. Yamagishi, “Privacy and utility of x-vector based speaker anonymization,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, to appear.

[3] B. M. L. Srivastava, N. Vauquier, M. Sahidullah, A. Bellet, M. Tommasi, and E. Vincent, “Evaluating voice conversion-based privacy protection against informed attackers,” in 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 2802–2806, 2020.

[4] N. Tomashenko, X. Wang, E. Vincent, J. Patino, B. M. L. Srivastava, P.-G. Noé, A. Nautsch, N. Evans, J. Yamagishi, B. O’Brien, A. Chanclu, J.-F. Bonastre, M. Todisco, and M. Maouche, “The VoicePrivacy 2020 Challenge: Results and findings,” Computer Speech and Language, vol. 74, pp. 101362, 2022.

[5] L. Girin, S. Leglaive, X. Bie, J. Diard, T. Hueber, and X. Alameda-Pineda, “Dynamical variational autoencoders: A comprehensive review,” Now Foundations and Trends, 2021.

[6] A. Baevski, H. Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” in Advances in Neural Information Processing Systems, pp. 12449–12460, 2020.

[7] B. M. L. Srivastava, A. Bellet, M. Tommasi, and E. Vincent, “Privacy-preserving adversarial representation learning in ASR: Reality or illusion?” in Interspeech, pp. 3700–3704, 2019.

[8] T. Ma, D. Li, W. Wang, and J. Dong, “CFA-Net: Controllable face anonymization network with identity representation manipulation,” arXiv preprint arXiv:2105.11137, 2021.

Profil du candidat :
Strong programming skills in Python/Pytorch.
Prior experience in speech and video processing will be an asset.

Formation et compétences requises :
MSc in computer science, machine learning, or signal processing.

Adresse d’emploi :