AI-based user authentication reinforcement by continuous extraction of behavioral interaction features

In this work, we conduct an experiment to analyze the feasibility of a continuous authentication method based on the monitorization of the users’ activity to verify their identities through specific user profiles modeled via Artificial Intelligence techniques. In order to conduct the experiment, a custom application was developed to gather user records in a guided scenario where some predefined actions must be completed. This dataset has been anonymized and will be available to the community. Additionally, a public dataset was also used for benchmarking purposes so that our techniques could be validated in a non-guided scenario. Such data were processed to extract a number of key features that could be used to train three different Artificial Intelligence techniques: Support Vector Machines, Multi-Layer Perceptrons, and a Deep Learning approach. These techniques demonstrated to perform well in both scenarios, being able to authenticate users in an effective manner. Finally, a rejection test was conducted, and a continuous authentication system was proposed and tested using weighted sliding windows, so that an impostor could be detected in a real environment when a legitimate user session is hijacked.