Building an accretive authentication system using a RBF network

A computerized authentication system should be able to admit new authentic entries continuously while maintain the existing entry records and an uninterrupted system operation. In this paper, we describe a competitive RBF neural network that is able to incrementally construct itself in response to the pattern samples presented to the system. The neural network is thus a suitable choice for authentication system applications. The accretion property of the neural network is made possible by allowing each pattern class (an authentic entry) being modeled in multiple hyper-ellipsoidal distributions, and mapping these distributions to multiple RBF neural units.