Adaptive Fusion of Biometric Scores by Fuzzy Weighting Based on Fuzzy C-Means
Aimé Kondalemba Lusasu *
Department of Mathematics and Computer Science, Faculty of Science and Technology, National Pedagogical University (UPN), Kinshasa, DRC.
Zéphirin N’teba Makala
Department of Mathematics and Computer Science, Faculty of Science and Technology, National Pedagogical University (UPN), Kinshasa, DRC.
José Indenge Y'esambalaka
Department of Mathematics and Computer Science, Faculty of Science and Technology, National Pedagogical University (UPN), Kinshasa, DRC.
Fidèle Mwaku Mvunzi
Department of Mathematics and Computer Science, Faculty of Science and Technology, National Pedagogical University (UPN), Kinshasa, DRC.
Alain Musesa Landa
Department of Mathematics and Computer Science, Faculty of Science and Technology, National Pedagogical University (UPN), Kinshasa, DRC.
Rostin Mabela Makengo
Department of Mathematics and Computer Science, Faculty of Science and Technology, National Pedagogical University (UPN), Kinshasa, DRC.
*Author to whom correspondence should be addressed.
Abstract
Multimodal biometric systems can improve recognition by combining information from multiple modalities, but the reliability of individual scores may vary across comparisons. This study evaluates Adaptive Fuzzy Weighting based on Fuzzy C-Means (AFW-FCM), a score-level fusion method in which the contribution of each modality is adjusted for each comparison according to confidence derived from the distribution of its scores. For each modality, a two-cluster Fuzzy C-Means model is learned from the training scores. The degree of membership of a score in the learned regions is used to define a confidence value, which is then normalised across the available modalities to obtain adaptive fusion weights. The method is evaluated primarily on BioSecure DS2 and additionally on NIST BSSR1 and a 50-subject subsample of CASIA-IrisV4. Under nominal conditions on BioSecure DS2, AFW-FCM achieved 97.4% accuracy, a 2.5% equal error rate, and an area under the curve of 0.995. On NIST BSSR1, it achieved 97.1% accuracy and a 2.8% equal error rate. On the CASIA-IrisV4 subsample, it achieved 96.1% accuracy, a 3.0% equal error rate, and an area under the curve of 0.985. Under Gaussian score perturbation at σ = 0.20, accuracy was 91.2%, compared with 82.8% for fixed weights. These results indicate that comparison-specific fuzzy weighting can maintain comparatively stable fusion performance under the controlled evaluation conditions used in this study.
Keywords: Multimodal biometrics, score fusion, Fuzzy C-Means, adaptive weighting, robustness