Recent studies have highlighted the susceptibility of Face Recognition Systems (FRS) to presentation attacks, also known as spoofing. Such attacks can enable malicious actors to gain unauthorized access to biometric systems. The current state-of-the-art spoofing detection methods have primarily concentrated on detecting 2D (photo and display video) and 3D-based presentation attacks. Despite limited focus on the feature extraction process, particularly regarding the generalization of learned representations, we introduce a fresh approach to face anti-spoofing. Our method emphasizes the feature extraction process, presenting a novel Adversarial Feature Learning framework. This framework aims to develop discriminative representations by learning features exclusively from genuine face samples. By initially focusing on learning features from bona fide presentations, we create a generalized model. Subsequently, by incorporating a small set of attack samples, we enhance the model’s ability to detect outliers, thereby addressing the challenge of identifying unforeseen attacks more effectively. The proposed novel approach is based on Adversarial Feature learning to identify printed mask or wrap attacks. The method utilizes the latent features of genuine facial samples learned from a trained VAE-GAN. Additionally, we have created a new 2D wrap attack dataset of 60 subjects. To validate the generalizability, we carried out detailed performance evaluation, in which the proposed approach outperforms the current state-of-the-art spoofing detection methods.