Multi-biometric Convolutional Neural Networks for Mobile User Authentication

Existing literature has established that multibiometric systems, which consolidate information from multiple biometric sources, can significantly enhance the performance by overcoming limitations such as non-universality, noisy sensor data and large intra-class variations. Several fixed and learning-based fusion methods exist to combine multibiometric information for personal verification. The aim of this paper is to develop convolutional neural networks (CNN) architectures for fusion of biometric information from multiple sources. The advantage of CNN-based multibiometric fusion include (a) the ability to perform early, intermediate and late fusion, and (b) the fusion architecture itself can be learned during network training. Experimental investigations on large scale VISOB dataset prove the efficacy of the multibiometric CNNs over conventional fusion methods.