Fuzzy few-nearest neighbor method with a few samples for personal authentication

The Fuzzy few-Nearest Neighbor (Ff-NN) method, which is an extended version of k-Nearest Neighbor algorithm (k-NN) and one of case-based learning methods, is proposed. Ff-NN intends to achieve stable identification performance even if the number of learning samples is as small as two. Applied to personal authentication systems such as enter/exit authorizations, Ff- NN reduces the user dictionary creation burden. Using 26 kinds of feature (face images and voices) data from 66 test objects, we conducted experiments on a PC to verify the feasibility of our proposed method. Forced recognition rate of conventional single-NN is 79.2% (standard deviation 2.83), and that of Ff-NN is 87.6% (SD 1.97). Recognition rates of dictionary data with 14, 17, and 26 features, are 90.6%, 92.5%, and 97.5%, respectively. We collect a very small number of nonintrusive samples so that two or more features are used to improve recognition performance. We present applicability of this method to personal authentication systems through experiments using 66 registrants, corresponding to 30 households.