Fingerprint Liveness Detection Project

In this project, we explored two different classification models (SVM & Neural Network) on four feature sets for fingerprint liveness detection.

About

We first extract features from fingerprint images and then develop machine learning models such as neural networks and SVM for classification of real and fake fingerprints. The feature extraction methods that we used were BSIF, WLD, LPQ, and CNN-RFW. When training our models, we implemented 10-fold cross-validation and some techniques to combat over-fitting such as dimensionality reduction of features and Gaussian noise layers. In order to have a larger training set, we also introduced new training data by augmenting available training images. We were able to achieve test accuracy ranging from 80% to 99% from different models, in which we outperform the contestants of the LivDet 2015 competition on Digital Persona data set.

See full text

Running the code

Please change input and output paths in the scripts asvappropriate and run in the order as numbered, bullet points mean scripts can be ran in any order.

Data set: We use the fingerprint images data set from "Liveness Detection Competitions 2015" Available Here.

1. Feature extraction

1. (optional, recommended) For the image augmentation, please refer to: augment_data.py
2. Four feature extraction methods applied:
   • Binary Statistical Image Filter(BSIF), in folder BSIF_feature
   • Local Phase Quantization(LPQ), in folder LPQ_feature
   • Weber Local Descriptor(WLD), in folder WLD_feature
   • Convnet Features: convnet_nn.py

2. SVM using sklearn in python

• Simple (Support Vector Machine (SVM) models) SVM_classifiy:
  1. Train and cross validate model: svm_fingerpint_liveness_detection_foldval.py is for cross validation
  2. Test model, and draw AUC and ROC curves: svm_fingerprint_liveness_detection_2015.py
• Convnet SVM models:
  1. Extract convnet features: CNN-RFW
     • Run "convnet_nn.py" with DO_TRAINING=True to generate the "pretrain.h5" model file, and then set DO_TRAINING=False to reuse previously trained model saved to this file
     • "cnnrandom" is required to extract features
  2. Train model: convnet_svm.py is for the CNN feature (result images are in svm_2015_result)

2. Neural Network using Keras in python

Simple Layer Neural Network models (models split based on what deatures they use):

• BSIF Features: 2015DigPer_test_BSIF.py

• LPQ Features: 2015DigPer_test_LPQ.py

• WLD Features: 2015DigPer_test_WLD.py

• Mix Features: 2015DigPer_test_MixFeat.py

• VGG16 Neural Network model:

  1. Train and test model: train_vgg16.py

• VGG16+vector multibranch Neural Neetwork model:

  1. Merge features: merge_features.py
  2. Train and test model: train_multi.py.

FYI: Their data generator functions can be found in gen.py

Summary of models used in this repository

Results