<p>Recently, biometric authentication has gained significant attention as a reliable method for personal information security. In this paper, we propose a photoplethysmography signal and a machine learning-based approach to develop a biometric authentication system. We employed a self-collected dataset comprising 55 PPG signals obtained through the BIOPAC MP36 system. The PPG signals were preprocessed by using wavelet transform and empirical mode decomposition to remove unwanted noise and to extricate salient features. Time and cepstral features, such as time auto-regressive, cepstral auto-regressive, heart rate variability, and cepstral activity, mobility, and complexity, were extracted from the three intrinsic mode functions. The fusion of these features resulted in the highest accuracy of 99.6% by using a K-nearest neighbor classifier. Finally, we utilized five types of feature selection and ranking algorithms to reduce the dimensionality of data while maintaining the same accuracy level. In contrast to others, ReliefF not only reported the same accuracy of 99.6% but also utilized 21% fewer features as compared to that of the full fusion approach. The findings highlight the significance of feature fusion for reliable and accurate biometric authentication systems, offering scalable solutions for AI-based cybersecurity frameworks and next-generation digital systems.</p>