Overfitting in multilayer perceptrons as a mechanism for (pseudo)random number generation in the design of secure electronic commerce systems

A novel methodology for generating strong pseudorandom numbers intended to be used in security mechanisms of communication systems for electronic commerce is presented. The proposed approach relies on the complex mapping properties of the multilayer perceptrons (MLP) to behave irregularly and unpredictably for a series of patterns chosen to be different from the ones encountered in their training procedure, when overfitting has occurred. In such a case the neural network attempts to fit perfectly the sample training data by involving a very complex surface and thus, its generalization capabilities are compromised. When, however, a neural network cannot generalize, if it is fed with unknown input patterns similar to the ones used during its training phase, then, its behavior is unpredictable since there is no analytic formula for determining the complex fitting surface of the training samples. The quality of these random number generators is investigated employing the most important statistical tests and it is found that is better than that of well, known in the literature random number generators.