Simplicity Wins: Lightweight Vision Transformer Heads for Presentation Attack Detection

The growing reliance on digital services has amplified the need for secure and seamless identity verification, particularly in regulation-sensitive domains such as online banking. Facial recognition-based Know Your Customer (KYC) procedures are increasingly targeted by presentation attacks (PAs), including printed photographs, video replays, and 3D masks. To mitigate these threats, Presentation Attack Detection (PAD) systems are critical components of robust authentication pipelines. While recent advances have centered around deep and complex neural network architectures to enhance PAD performance, these models often exhibit limitations when faced with out-of-distribution samples and scarce training data.In this study, we explore a data-centric approach to PAD by leveraging a pre-trained Vision Transformer (DINOv2) as a fixed feature extractor, coupled with a lightweight classification head. We hypothesize that, given access to sufficiently diverse training data, such a simplified architecture can match or exceed the performance of more intricate models. Experimental results across multiple attack scenarios validate our hypothesis, demonstrating competitive accuracy and improved generalizability. These findings emphasize the efficacy of scalable, data-driven methodologies over architectural complexity in PAD system design.