Projet de recherche doctoral numero :3028

Date depot: 1 janvier 1900
Titre: Une nouvelle approche co-modale en biométrie
Directeur de thèse: Jean-Luc DUGELAY (Eurecom)
Directeur de thèse: Nicholas EVANS (Eurecom)
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Non defini

Resumé: The goal of this thesis is to develop the first co-modal approach to biometrics using the co-training of two biometric classifiers. The first 12-18 months will be dedicated to the understanding and implementation of a co-training framework. There will be three aspects to this work: • Co-training – here the aim will be to establish the fundamental co-training framework with which to undertake the remainder of the proposed research. This element of the work will compare the performance of existing standard uni-modal and multi-modal classifiers to co-trained, co-modal classifiers using semi-supervised learning. • Data quantities – it will be necessary to investigate the effect on performance of differing amounts of labeled data in order to determine how much is required for good performance. Similarly, different sized pools of unlabelled data will provide a greater spread of samples from which to choose new data for co-training. Both the size of the pool and different approaches to confidence estimation (i.e. the approach used to choose unlabeled data for co-training) will be investigated. • Convergence and efficiency – co-training requires the iterative training of the two different classifiers and is therefore likely to be computationally expensive. Reduced computational load can be achieved by increasing the amount of data used during each iteration of co-training, thereby reducing the number of iterations required to exhaust the supply of un-labelled data. However, this necessarily involves the repeated use of data with which the other classifier has less confidence and so, ultimately, such an approach is likely to degrade performance as errors are made in automatic classification. Therefore it will be necessary to evaluate the trade-off between the rate of training and classification performance. The work will require a significant theoretical study of co-training algorithms in addition to the practical work involved in implementing the co-training framework and undertaking extensive related experimentation. Existing open-source and in-house state-of-the-art speaker and face recognition systems will be employed, thereby removing the burden of developing and optimising individual classifiers. In addition, pre-extracted speech and face features will be provided so that the focus during the first period will be entirely on co-modal biometrics and not on the design or implementation of individual classifiers or biometrics. Nonetheless it will be necessary for the PhD student to familiarise themselves with the training or adaptation of the two classifiers as new training data is identified. Supervision from both academic advisors and current post-doctoral research engineers will be provided to ensure the successful implementation of a co-training system. However, the above work requires the implementation from scratch of a new co-modal approach to biometrics. This is challenging, involved work and with the other aspects described above, it is expected to take between 12 and 18 months to complete. Results include the world’s first implementation and study of a co-modal biometric system. The co-training aspects of the study are entirely novel within the biometrics domain and will lead to publications in the most respected biometrics conferences and journals. Given the novelty and potential of the subject, these should realistically come early in the PhD study, during the first 12-18 month period and will provide an appropriate performance indicator for the student’s mid-term PhD evaluation. Given the adventurous nature of the proposed research, the second 18 months of the work must remain suitably flexible. The first review and the end of the first 12 months will produce an initial plan for the final 18 months but will be periodically reviewed every 6 months in the second half of the PhD programme. Topics to be considered include optimization of the co-modal biometric system by considering different features and classifiers and the possibility of utilizing co-modal biometrics for improved robustness to spoofing or forgery attacks from dedicated impostors: • Features and classifiers – it will be necessary to determine which features and individual classifiers lead to the best performance when combined in a co-modal approach. Co-training relies upon the two views of the data (here biometric modes) being independent, i.e. so that each classifier has the potential to correct mistakes made by the other, and so it will be necessary to identify both features and classifiers which complement each other well when employed for co-training. • Spoofing – since the co-training approach proposed here employs a classifier based on one biometric trait to train another classifier on a second, different biometric trait, co-modal training may have potential to protect systems from spoofing attacks. The topic of biometric spoofing is certainly not new, and there are some recent p

Doctorant.e: Zhao Xuran