Score-based fusion of ASV (ISV-based) and PAD systems

To edit a toolchain, please use a modern browser (Mozilla Firefox 3.0+, Google Chrome 1+, Apple Safari 3+, Opera 9.5+, Microsoft Internet Explorer 9+)

This toolchain implements an ASV-PAD integrated system for speech data. Integration is done based on score fusion, which, in turn, assumes a presence of a classifier (e.g., Logistic Regression). ASV toolchain implements speaker verification using Intersession Variability Modelling (McCool2013]) described in [McCool2013]:

The fusion toolchain consists of three parts: 1. ASV system ASV system is evaluated on both zero-impostors and on spoofing attacks. Training set is used for both training UBM and for computing the scores that are used to train fusion classifeir later.

Hence, ISV-basd modelling is as follow: The extracted features from the training set are used to train the Universal Background Model (UBM). The algorithm GMM can be used for this purpose. For each set of feature vectors of one image, the GMM statistics are extracted (based on the Maximum a Posteriori (MAP) adaption using the UBM as a prior). The algorithm GMM Statistics can be used for this purpose. The GMM Statistics of the training set are used to estimate the U subspace. The algorithm ISV can be used for this purpose. For each set of GMM Statistics of a given client, the UBM and the U subspace, a client can be enrolled using the algorithm ISV Enroll The scoring step for the ISV is defined as the LLR between the client model and the UBM, using the GMM Statistics of a given probe as input. The algorithm ISV Scoring can be used for this purpuse. The main inputs for this algoritm are: the GMM Statistics of a probe, the client model, the UBM, the U subspace.

2. PAD system Here, training set is also used for both training the PAD classifier and for computing the scores that are used to train fusion classifeir later.

3. Learning-based score fusion Classifier is trained on the scores computed on training sets of ASV and PAD systems. Positive training set consists of real scores from PAD and correct identities scores from ASV. Negative training set consists of attack scores from PAD and zero-impostor scores from ASV. Then, fused scores from dev and test sets (from both ASV and PAD systems) are computed using this classifer, e.g, a pair of (ASV, PAD) scores is projected on the classifer's model and a fused single score is obtained.

[McCool2013]
  1. McCool, et al.: Session variability modelling for face authentication. IET biometrics 2.3 (2013): 117-129.
Updated Name Databases/Protocols Analyzers
pkorshunov/pkorshunov/isv-asv-pad-fusion-complete/1/asv_isv-pad_lbp_hist_ratios_lr-fusion_lr-pa_aligned avspoof/2@physicalaccess_verify_train,avspoof/2@physicalaccess_verification,avspoof/2@physicalaccess_verification_spoof,avspoof/2@physicalaccess_verify_train_spoof,avspoof/2@physicalaccess_antispoofing pkorshunov/spoof-score-fusion-roc_hist/1
pkorshunov/pkorshunov/isv-asv-pad-fusion-complete/1/asv_isv-pad_gmm-fusion_lr-pa avspoof/2@physicalaccess_verify_train,avspoof/2@physicalaccess_verification,avspoof/2@physicalaccess_verification_spoof,avspoof/2@physicalaccess_verify_train_spoof,avspoof/2@physicalaccess_antispoofing pkorshunov/spoof-score-fusion-roc_hist/1
Terms of Service | Contact Information | BEAT platform version 2.2.1b0 | © Idiap Research Institute - 2013-2024