Deep Learning Based Sequential Mining for User Authentication in Web Applications

Abstract

Behavioral biometrics is a seamless and transparent way to authenticate or identify users through their interaction with electronic systems. It can serve as an additional security mechanism to existing security methods by continuously authenticating the users, for example when using pointing devices (e.g., mouse, touchscreen). These methods usually aim at extracting meaningful features such as curvature and acceleration using the raw mouse coordinates and ignore the specific elements the user interacts with during the movement. A possible improvement is to combine these methods with approaches that analyze the user path of elements throughout the session. One such previously suggested process proposes using a model-per-user approach, built using the traditional sequence mining algorithm Hidden Markov Model (HMM). In this paper we examine the use of deep learning sequential mining mechanisms for authentication, using mechanisms such as Long Short-Term Memory (LSTM), LSTM with Attention, and a Convolutional Neural Network (CNN). This method has the major advantage of one global model per web application, which drastically reduces the system’s required memory and storage resources. We demonstrate the competitive advantage by encouraging results in low false positive rates (FPR) ranges on an anonymized dataset collected by IBM from accounts of more than 2000 web application users.