Image-Driven Face Presentation Attack Detection in Biometric Systems

As biometric systems become increasingly relied on secure authentication, concerns about their susceptibility to Face Presentation Attacks (FPAs) have intensified. This paper introduces a framework for detection of image-driven FPAs in biometric systems, focusing on enhancing the security of face recognition technologies. We propose a methodology that integrates Histogram of Oriented Gradients (HOG) with deep learning techniques. This approach utilizes pretrained models and custom layers to effectively differentiate between legitimate user images and various forms of presentation attacks, including those generated by GANs, morphing techniques, and photoshop manipulations. Through extensive experimentation, we analyzed the performance of each model, offering detailed comparisons concerning accuracy, precision, and F1 score across various categories as well as overall performance of each model. Our findings imply that fusion of HOG with a fine-tuned VGG16 model can achieve significantly higher overall accuracy.