Intelligent Systems Research Lab

Self-Supervised Federated Learning for Face Presentation Attack Detection

TUBITAK 1002 Project (2024 – 2025)

Principal Investigator Mehmet Fatih Gündoğar

Consultants: Prof. Çiğdem Eroğlu Erdem, Assoc. Prof. Ömer Korçak

Abstract: Accurate and reliable operation of facial recognition and verification systems relies heavily on the detection of presentation attacks. Spoofing materials utilized for face presentation attacks include printed images, high-resolution video recordings taken with mobile devices, and realistic-looking masks. The proliferation of diverse and sophisticated spoofing methods employed in presentation attacks poses significant challenges for presentation attack detection (PAD). Many different kinds of data sets have been incorporated into the literature, serving as valuable resources for researchers who are dealing with PAD methods. However, the available datasets have limitations in terms of the types of presentation attacks they cover, and it is often the case that a solution that Works well for one dataset may not be applicable to other datasets. The success rate of machine learning and deep learning models drops by quite a bit when they are tested on a dataset taken under different conditions after being trained with a small set of data from a few datasets. This is because of the domain shift issue, which means that the applicable test dataset may include a variety of attack types, data gathered in varying illumination and resolution quality settings, and subjects with varying ethnic backgrounds. To enhance the efficacy of deep learning models, one potential approach is augmenting the quantity of data required through sourcing it from various sources and data centers. This kind of data acquisition is extremely difficult since it needs a lot of data to be labeled, the personal information of the subjects needs to be protected, and the data to be collected in a single, centralized data center. To overcome these issues, this project aims to present a self-supervised federated learning PAD framework. The federated learning approach eliminates the need for collecting data on a central server. Besides, the suggested method allows the use of unlabeled local data in the self-supervised task-learning phases, which improves the training effectiveness of local models and the global model stored on the central server. To the best of our knowledge, there is currently no documented PAD method that utilizes a self-supervised federated learning (FL) approach in the literature.