Anti-spoofing few-shot learning model for face recognition

• Main Author: Ang, Ting Feng
• Other Authors: Ong Chin Ann
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Few-shot; Anti-spoofing; Machine learning
• Online Access: https://hdl.handle.net/10356/181185
• Institution: Nanyang Technological University

Face recognition technology has been widely used in a variety of industries recently, including mobile devices and security systems. However, a serious security risk arises from these systems' vulnerability to spoofing attacks, which use images, videos, or 3D masks to trick recognition algorithms. The goal of this study is to strengthen the resilience of face recognition systems against these types of attacks by developing an Anti-Spoofing Few-Shot Learning Model. The suggested model uses few-shot learning, a machine learning technique, to help it perform effectively even with a few training samples. Few-shot learning makes the model more robust and versatile in real-world applications by enabling it to adjust to novel spoofing tactics with little data quickly. In order to distinguish between real and fake inputs, the model is trained on a variety of datasets that include both spoof and real face photos. This study investigates cutting-edge machine learning models for recognizing faces with few examples, emphasizing advanced techniques like Prototypical Networks, Matching Networks, and Relation Networks. Every model has its own specific advantages when dealing with the task of facial recognition using very little training data. Prototypical Networks streamline classification with a prototype for each class, improving the model's capacity to generalize from limited examples. Matching Networks use an attention mechanism to compare a query image with a limited number of labeled examples, which enhances their ability to recognize new identities. Relation Networks improve this process by developing skills to assess and measure the connection between image pairs, allowing for reliable recognition even in difficult situations with significant variation within the same class. By thoroughly assessing these models, this study seeks to promote the progress of dependable and effective few-shot face recognition systems.