Quad-Biometrics for Few-Shot User Identification

Recent research in AI-based user identification using biometrics is actively underway. User identification based on biometrics is widely used in various industries and energy sectors because it is challenging to counterfeit or tamper with. AI models extract features from a user's bio information and perform identification. However, AI models are vulnerable to adversarial attacks that exploit weight information, resulting in misclassification and reduced reliability, particularly for specific biometrics. Adequate training data is also essential. In this paper, we propose a multi-biometric-based user identification system that is robust against adversarial attacks and effective with limited training data. We evaluate user identification performance based on the multi-modal fusion method. Our proposed method consists of feature extraction from biometrics and fusion for multi-biometric user identification. When evaluating user identification performance using data collected from CASIA-BIT, including face, iris, palmprint, and fingerprint data, our proposed method demonstrates superior accuracy of 24% and an F1-score of 0.3033 compared to single-biometric approaches. Furthermore, when using quad-biometrics, our method improved accuracy of 31% and an F1-score of 0.3717 and an F1-score of 0.3717 compared to dual-biometrics, with an 88.33% faster training speed.