Score files to reproduce the results of the chapter

Due to the large number and size of score files, some of them need to be downloaded and the rest were archived and need to be unzipped. Please follow these steps to prepare the scores, so that error rates and figures used in the chapter can be computed:

• Unzip pad_handcrafted.zip and pad_megafusion.zip files locate in the scores’ folder into respective `pad_handcrafted and pad_megafusion subfolders.

• Download and unarchive some of the scores of handcrafted features-based PAD systems:

  $ cd scores $ wget http://www.idiap.ch/resource/biometric/data/interspeech_2016.tar.gz $ tar -xzvf interspeech_2016.tar.gz $ Add these score files inside pad_handcrafted folder

• Download and unarchive scores for CNN-based PAD systems:

  $ cd scores $ wget http://www.idiap.ch/resource/biometric/data/isba2018-pad-dnn.tar.gz $ tar -xzvf isba2018-pad-dnn.tar.gz $ Rename the obtained scores folder into pad_cnn folder

Reproducing the results of the chapter

To compute error rates presented in Tables 5 and 6 of the chapter on performance of PAD systems based on handcrafted features, the following script should be used:

$ ./compute_handcrafted_results.sh

The script will create a folder for each configuration of PAD system and datasets. The folders contain DET curves for each given configuration, histograms of score distributions, and error rates. The script also creates a text file ‘latex_table_handcrafed_stats_eer.txt’ which contains a LaTeX formatted table with the results from Table 6 and 6.

To compute error rates presented in Table 7, run the script that is installed with bob.measure using the following:

$ bob_eval_threshold.py -c eer ../scores/pad_megafusion/corresponding_folder/scores-dev

To compute error rates presented in Tables 10 and 11 and Figure 6 of the chapter on performance of CNN-based PAD systems, the following script should be used:

$ ./compute_cnn_results.sh

The script will create a folder for each configuration of PAD system and datasets. The folders contain DET curves for each given configuration (subset of these plots are shown in Figure 6 of the chapter), histograms of score distributions, and error rates. The script also creates a text file ‘latex_table_cnn_stats_eer.txt’ which contains a LaTeX formatted table with the results from Table 10 and 11.

To compute error rates for different types of attack as presented in Table 12, please execute the following script:

$ ./compute_results_cnn_per_attack.sh

The script will compute error rates for each type of attack of voicePA and BioCPqD-PA databases used in the paper and produce ‘latex_table_cnn_per_attack_stats_eer.txt’ file with LaTeX formatted table that has results from Table 12.