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

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