LAMARCKIAN EVOLUTION OF NEURAL NETWORKS APPLIED TO KEYSTROKE DYNAMICS

The pace of computing and communications development has contributed to an increased data exposure and, consequently, to the rise of an issue known as identity theft. By applying user profiling, which analyzes the user behavior in order to perform a continuous authentication, protection of digital identities can be enhanced. Among the possible features to be analyzed, this paper focuses on keystroke dynamics, something that cannot be easily stolen. As keystroke dynamics involves dealing with noisy data, it was chosen a neural network to perform the pattern recognition task. However, traditional neural network training algorithms are bound to get trapped in local minimum, reducing the learning ability. This work draws a comparison between backpropagation and two hybrid approaches based on evolutionary training, for the task of keystroke dynamics. Differently from most evolutionary algorithms based on Darwinism, this work also studies Lamarckian evolutionary algorithms that, although not being biologically plausible, attained promising results in the tests.