Graduate School and Research Center in Digital Sciences

Privacy-preserving speaker recognition with cohort score normalisation

Nautsch, Andreas; Patino, Jose; Treiber, Amos; Stafylakis, Themos; Mizera, Petr; Todisco, Massimiliano; Schneider, Thomas; Evans, Nicholas

INTERSPEECH 2019, 20th Annual Conference of the International Speech Communication Association, September 15-19, 2019, Graz, Austria

In many voice biometrics applications there is a requirement to preserve privacy, not least because of the recently enforced General Data Protection Regulation (GDPR). Though progress in bringing privacy preservation to voice biometrics is lagging behind developments in other biometrics communities, recent years have seen rapid progress, with secure computation mechanisms such as homomorphic encryption being applied successfully to speaker recognition. Even so, the computational overhead incurred by processing speech data in the encrypted domain is substantial. While still tolerable for single biometric comparisons, most state-of-the-art systems perform some form of cohort-based score normalisation, requiring many thousands of biometric comparisons. The computational overhead is then prohibitive, meaning that one must accept either degraded performance (no score normalisation) or potential for privacy violations. This paper proposes the first computationally feasible approach to privacy-preserving cohort score normalisation. Our solution is a cohort pruning scheme based on secure multi-party computation which enables privacy-preserving score normalisation using probabilistic linear discriminant analysis (PLDA) comparisons. The solution operates upon binary voice representations. While the binarisation is lossy in biometric rank-1 performance, it supports computationally-feasible biometric rank-n comparisons in the encrypted domain.

Document Bibtex

Title:Privacy-preserving speaker recognition with cohort score normalisation
Keywords:privacy, speaker recognition, score normalisation, binary keys, secure computation
Type:Conference
Language:English
City:Graz
Country:AUSTRIA
Date:
Department:Digital Security
Eurecom ref:5937
Copyright: © ISCA. Personal use of this material is permitted. The definitive version of this paper was published in INTERSPEECH 2019, 20th Annual Conference of the International Speech Communication Association, September 15-19, 2019, Graz, Austria and is available at :
Bibtex: @inproceedings{EURECOM+5937, year = {2019}, title = {{P}rivacy-preserving speaker recognition with cohort score normalisation}, author = {{N}autsch, {A}ndreas and {P}atino, {J}ose and {T}reiber, {A}mos and {S}tafylakis, {T}hemos and {M}izera, {P}etr and {T}odisco, {M}assimiliano and {S}chneider, {T}homas and {E}vans, {N}icholas}, booktitle = {{INTERSPEECH} 2019, 20th {A}nnual {C}onference of the {I}nternational {S}peech {C}ommunication {A}ssociation, {S}eptember 15-19, 2019, {G}raz, {A}ustria}, address = {{G}raz, {AUSTRIA}}, month = {09}, url = {http://www.eurecom.fr/publication/5937} }
See also: