Face anti-spoofing by using diffusion model

Abstract

This thesis explores the potential of diffusion models for improving face anti-spoofing (FAS) techniques. Due to the increasing usage of facial recognition systems in various domains, such as smartphone unlocking, airport security, and financial transactions, the need for robust FAS techniques has escalated. Existing FAS datasets, such as OULU-NPU, are limited in size and diversity, making developing effective FAS techniques difficult. To overcome these challenges, this study utilizes generative models such as generative adversarial networks, denoising diffusion probabilistic models, and stable diffusion to expand FAS datasets. The primary objectives are to implement various generative models to create spoof facial image data for FAS training and validate the improved performance of FAS methods with these expanded datasets.