Presentation Attack Detection for Multispectral Face Biometric System Using Federated Learning

These days, most biometric devices use multi-spectral band images to improve performance and avoid both weather and illumination limitations. Developing a multi-spectral biometric model requires a large amount of face biometric data and it is challenging because of the sensitive nature due to privacy concerns of biometric data. Presentation Attacks are the major spoofing challenges in biometric systems that use printouts or videos of biometric data. Federated learning (FL), is a collaborative machine-learning technique for sensitive data such as medical images which is also the best solution for developing biometric models in a collaborative training method. One of the major challenges FL technique is model aggregation due to the heterogeneous data among the members. Presently there are many techniques available for the heterogeneous data aggregations for both centralized and decentralized FL. Here we are using the FL technique to develop a Presentation Attack Detection (PAD) system multispectral face biometrics by preserving the privacy of the data. These data are heterogeneous and use different model aggregation techniques both centralized and decentralized. In this work, we are proposing the comparative study of different techniques for a FL-based multispectral face biometric with a PAD system.