Improved Face Morphing Attack Detection Method Using PCA and Convolutional Neural Network

Face recognition is the most extensively utilized security and public safety verification method. In many nations, Automatic Border Control system uses face to confirm identification of travelers The ABC vulnerable morphing attacks; systems give acceptance for traveller, even though passport photo does not represent actual image person but a result merger two images. Therefore, it vital determine whether altering (morph) or actual. This research proposes an improved method extract features from facial proposed consists four phases: first stage, morph images were generated using set databases real people, used every that similar in general shape landmarks producing morphed three types techniques this field (Automatic selection landmark, StyleGAN, Manual landmark). StyleGAN has been relied upon achieve best results artefact-free second phase, Faster Region Convolution neural network utilizing determining cutting important area (eyes, nose, mouth, skin) face, where we leave hair, ears, background database. third are extracted Principal component analysis, eigenvalue, eigenvector; matrix two-dimensional with one layer each technique. Then merge (with out s) into layers. represents principal analysis features, eigenvalue eigenvector features. Finally, introduced convolutional networks obtain optimal fourth phase classification process Deep Neural Network (DNN) classifier Support Vector Machine (SVM) classifier. DNN achieved average accuracy 99.02% compared SVM, 98.64%. power work evident through FRA RFF evaluation. Which values as low possible FAR 0.018, indicating error rate calculating actual, FRR 0.003, meaning morphed, 0.023, 0.06 SVM whenever these ratios less than one, higher system's detection. AMSL dataset (Accuracy 95.8%, 0.039, 0%) 95.2%, 0.047, 0.98) respectively. It turned training optimized significantly affects finding difference discovering modified images, case minor modifications dataset.