A New Learning Algorithm Based on Feedforward Neural Networks: Application in User Authentication

This paper introduce a new type of Neural Networks which although the class of training patterns is specified (supervised training) these classes is notused as desired output and the network will find them indirectly. So training algorithm, in addition to weights must specify desired outputs. We call this algorithmUnspecified Desired Output Supervised (UDOS) learning algorithm. To find the network weights, the extended MSE criterion is developed which measuresquantities of both class tightness and class separation. Despite the complexity of this multi-objective optimization task, the results of this study are clear.UDOS succeeded in achieving superior performance than the conventional methods. It is applied to thespecific practical problem of keystroke identification, with success.

Keywords:

Training algorithm , Quasi-Newton optimization , extended MSE , keystroke dynamics , within-class and between-class scatter matrices