Behavioral embedding for continuous user verification in global settings

Abstract

In recent years, user identity verification techniques based on mobile touch interaction are becoming more reliable and prominent. These techniques can be integrated in many types of mobile applications and help preventing illegitimate access to information and storage done by impostors. Such techniques usually rely on binary classification algorithms building models for each user in the system, that require other users data in order to build ones model and in case the other users’ data does not exists, a model cannot be built.In addition using a method that generates a model for each user in the system pose a significant challenge when scaling up to real-world systems. The main drawback with having a great number of models resides in the fact that it is difficult to debug, analyze, update, and fine-tune each individual model. Thus, we introduce a method for generating a unified global model that can verify the identity of every user in the system. Our most fundamental challenge is to preserve the unique behavior of each user in one unified model. Having a global model that was built prematurely also has privacy preserving characteristics, when compared to methods that require other users data. The core idea of our method is a novel behavioral embedding layer that captures and embeds each user’s unique behavior and enables it to be used within global settings. We compared our method to several state-of-the-art techniques in experiments with over 9,000 users on 85 different devices. Our method achieves 0.918 AUC and 15.6% EER in efficient global settings, bypassing the second-best method by a large margin of 0.119 on AUC and 11.7% on EER.