Person-Specific-Face-Anti-Spoofing

Introduction

This repo provides the code for our TIFS paper "Person Specific Face Anti-Spoofing with Subject Domain Adaptation". Different from previous works, we proposed to perform face anti-spoofing in person-specific manner, which means each subject is equipped with a unique classifier for distinguishing genuine and fake face images. Moreover, we considered a challenging case that some subjects' fake face images are not available. Toward such a problem, we proposed a subject domain adaptation algorithm to generate virtual samples for training. The experimental results proves the superiority of person-specific face anti-spoofing method compared with generic face anti-spoofing methods. Also, it is shown the proposed subject domain adaptation algorithm can generate virtual samples to assist the training of person-specific classifiers.

Experiments

Preparation

• libsvm: this open library can be found here.
• features: the off-the-shelf features extracted for this paper can be downloaded from my Baidu Driver here or Google Driver here

Run the code

Use Replay-Attack dataset as the example. Go to REPLAY-ATTACK folder, and then:

1. Load Feats and Labels.

   We need load all features and labels for Replay-Attack dataset. Run

   load('Feats_Replay-Attack_tr_fall_soriginal.mat')  % load train set features
   load('Labels_Replay-Attack_tr_fall_soriginal.mat') % load train set labels
   load('Feats_Replay-Attack_dev_fall_soriginal.mat') % load development set features
   load('Labels_Replay-Attack_dev_fall_soriginal.mat') % load development set labels
   load('Feats_Replay-Attack_te_fall_soriginal.mat') % load test set features
   load('Labels_Replay-Attack_te_fall_soriginal.mat') % load test set labels
   load('Feats_Replay-Attack_en_fall_soriginal.mat') % load enrollment set features
   load('Labels_Replay-Attack_en_fall_soriginal.mat') % load enrollment set labels

2. Estimate Transformation.

   Go to folder Estimate_Transformationn_for_Expe, fund runall.m:

   Feat_Types = [1 3];
   BLabels = {'ContN' 'AdvN'};
   
   for t = 1:length(Feat_Types)
       for b = 1:length(BLabels)
           EstimateTransformation_Enroll_Iterative(Feats_enroll_SL, Labels_train_SL, Labels_devel_SL, Labels_test_SL, Labels_enroll_SL, BLabels{b}, Feat_Types(t), 'CS');
           EstimateTransformation_Enroll_Iterative(Feats_enroll_SL, Labels_train_SL, Labels_devel_SL, Labels_test_SL, Labels_enroll_SL, BLabels{b}, Feat_Types(t), 'OLS');
           EstimateTransformation_Enroll_Iterative(Feats_enroll_SL, Labels_train_SL, Labels_devel_SL, Labels_test_SL, Labels_enroll_SL, BLabels{b}, Feat_Types(t), 'PLS');
           EstimateTransformation_Enroll_PCA(Feats_enroll_SL, Labels_train_SL, Labels_devel_SL, Labels_test_SL, Labels_enroll_SL, Feat_Types(t), BLabels{b}, 'PCA');  
           % EstimateTransformation_Enroll_PCA_withSubNum(Feats_enroll_SL, Labels_train_SL, Labels_devel_SL, Labels_test_SL, Labels_enroll_SL, Feat_Types(t), BLabels{b}, 'PCA');
       end
   end	

   In this file, we provided all methods to estimate the transformations, including CS, OLS, PLS as stated in the paper. We also provided a PCA-based transformation estimation. To get the transformation, comment the line as you need. Once finished, you will obtain a mat file with prefix "transform_".

3. Domain Adaptation. After get the estimated transformation, we can perform domain adaptation between subjects. Go to Domain_Adaptation folder, and find runall.m:

   Feat_Types = [1 3];
   Methods = {'CS' 'OLS' 'PLS' 'PCA'};
   
   for t = 1:2
       for m = 1:4 % length(Methods)
           TargetDA_AllQualities(Feats_train_SL, Labels_train_SL, Feats_devel_SL, Labels_devel_SL, Feats_test_SL, Labels_test_SL, Feats_enroll_SL, Labels_enroll_SL, Feat_Types(t), Methods{m})        
       end
   end

   Modify the variable Methods accordingly to do domain adaptation for different methods. Afterward, you will find mat files with prefix "SynthFeatures_".

4. Train Model.

   Congratulate to see you here. You are almost set! Go inside Train_Models folder. You can train generic face anti-spoofing model using runall_gen.m, while person-specific face anti-spoofing model using runall_ps.m. Meanwhile, you can also use runall.m to train models with the number of subjects used for training.

5. Test Model

   Finally, you may want to evaluate the performance. Go to Test_Models folder, find runall_ps.m and runall_gen.m and run them.

Citation

If you find our paper or code is useful in your researches, please consider citing:

@article{yang2015person,
  title={Person-specific face antispoofing with subject domain adaptation},
  author={Yang, Jianwei and Lei, Zhen and Yi, Dong and Li, Stan Z},
  journal={IEEE Transactions on Information Forensics and Security},
  volume={10},
  number={4},
  pages={797--809},
  year={2015},
  publisher={IEEE}
}