Biometric template protection of deep templates

The Idiap Research Institute seeks qualified candidates for a PhD position on "biometric template protection of deep templates".

Current research into applying Deep Neural Networks (DNNs) to face and speaker biometric recognition systems shows superior accuracy compared to traditional approaches, so it seems reasonable to assume that future state-of-the-art face and speaker recognition systems will be primarily DNN-based.

Recent advances in DNN-based face and speaker biometric systems have shown a preference for training the DNNs on raw face and speaker data as opposed to using pre-extracted features. While this allows for a greater level of automation in the biometric system, the added convenience may come at the risk of increased ignorance about the inner workings of the system. Consequently, it may become more difficult to predict what features of each biometric trait the system will use for recognition purposes. This introduces a new challenge to the development of suitable biometric template protection techniques to preserve the privacy of our biometric data.

The aim of this project will be to investigate the application of biometric template protection to DNN-based face and speaker recognition systems. The applicant will explore the suitability of current methods and homomorphic encryption applied to binarised ”embeddings” and more particularly the following results are expected:

  • Study of the types of embeddings produced by DNN-based face and speaker recognition systems;
  • Investigation into the suitability of current template protection schemes (e.g., Bloom filters and homomorphic encryption) for the DNN-based systems;
  • Proposal of new directions in biometric template protection to better take into account the nature of DNN-based biometric systems;
  • Study/development of techniques to evaluate the security and privacy preserving properties of the protected deep templates (e.g., relative entropy);
  • Open-source implementations of the investigated (and perhaps new) biometric template protection schemes on face and speaker recognition DNN baselines.

This project will be funded by the H2020 Marie Sklodowska-Curie program and the applicant will also be expected to be involved in project meetings, training activities as well as visiting research groups and/or companies.

The proposed research will be carried out in the Biometric Security and Privacy group (Dr. Sebastien Marcel) at Idiap. The research will rely on previous knowledge and softwares developed at Idiap, more specifically Bob toolkit.
Reproducible research is a cornerstone of the project. Hence a strong involvement in open source libraries such as Bob are expected.

The ideal candidate should hold a Master degree in computer science, electrical engineering or related fields. She or he should have a background in statistics, applied mathematics, optimization, linear algebra and signal processing. The applicant should also have strong programming skills and be familiar with Python, C/C++ (MATLAB is not a plus), various scripting languages and with the Linux environment. Knowledge in machine learning and more particularly deep learning (TensorFlow and pyTorch) is an asset. Shortlisted candidate may undergo a series of tests including technical reading and writing in English and programming (in Python and/or C/C++).

Idiap is affiliated with Ecole Polytechnique Federale de Lausanne (EPFL). Working at Idiap in Martigny, the successful candidate will become a doctoral student at EPFL, and thus also has to be accepted for enrolment by the Electrical Engineering (EDEE) or the Computer, Communication and Information Sciences (EDIC) Doctoral Schools. The PhD position is for 4 years, provided successful progress, and should lead to a dissertation. Annual gross salary ranges from 47,000 CHF (first year) to 50,000 CHF (last year).

Prospective candidates should apply through the Idiap Online Recruitment System (ORS) only, and applications should contain a cover letter (max 2 pages long and should explain how your background fit with our research activities -- see the short description of our research and our latest publications), CV, statement of research interests and accomplishments, and names and email addresses of 3 references. In your CV, relative ranking should be included when possible, such as average grade or ranks in the class. Inquiries about the position can be addressed to Dr. Sébastien Marcel.

The project start date is expected to be late 2019 or early 2020. The position will remain open until filled.

Interested? Please submit your candidature through the Idiap online recruitment system:

Biometric template protection of deep templates