SHRIMPS: A framework for evaluating multi-user, multi-modal implicit authentication systems

Abstract

Smart devices are commonly used in multi-user scenarios, such as shared household devices and shared corporate devices for front-line workers. A multi-user device requires both identification and authentication to defend against unauthorized access and distinguish between legitimate users in real-time, especially when multiple users participate in the same session. Although implicit authentication (IA) has been proposed to provide continuous and transparent authentication throughout a session, most existing IA solutions are optimized for single-user scenarios. The challenges of designing multi-user IA systems include fusing multiple modalities for good accuracy, segmenting and labeling behavioral data while authenticating, and adapting IA models to new users and new incoming data. We propose SHRIMPS, an evaluation framework to support IA researchers in the design of multi-user, multi-modal IA systems. SHRIMPS allows the evaluation of multi-user IA solutions that incorporate multiple modalities and supports adding new users and automatically labeling new incoming data for model updating. SHRIMPS supports different score fusion strategies, including a novel score fusion strategy based on Dempster-Shafer (D-S) theory to improve accuracy with considering uncertainties among different IA mechanisms. SHRIMPS enables composing tasks with public datasets to evaluate and compare different IA schemes. We present and evaluate two sample use cases to showcase how SHRIMPS helps address practical design questions of multi-user, multi-modal IA systems. The evaluation results show that D-S theory based score fusion methods can effectively reject attackers and detect user switches for the multi-user scenario in real-time.