Seamless authentication is a desired feature which is becoming more and more relevant, due to the distribution of personal and wearable mobile devices. With seamless authentication, biometric features such as human gait, become a way to control authorized access on mobile devices, without actually requiring user interaction. However, this analysis is a challenging task, prone to errors, with the need to dynamic adapt to new conditions and requirements, brought by the dynamic change of biometric parameters. In this paper we present a novel deep-learning based framework for gait-based authentication. The paper presents an in depth study of the building and training of a Recurrent Convolutional Neural Network with a real dataset based on gait reading performed through five body sensors. We introduce methodologies to further increase the classification accuracy based on data augmentation and selective filtering. Finally we will present a complete experimental evaluation performed on more than 150 different identities.