mFingerprint: Privacy-Preserving User Modeling with Multimodal Mobile Device Footprints

Abstract

Mobile devices collect a variety of information about their environments, recording “digital footprints” about the locations and activities of their human owners. These footprints come from physical sensors such as GPS, WiFi, and Bluetooth, as well as social behavior logs like phone calls, application usage, etc. Existing studies analyze mobile device footprints to infer daily activities like driving/running/walking, etc. and social contexts such as personality traits and emotional states. In this paper, we propose a different approach that uses multimodal mobile sensor and log data to build a novel user modeling framework called mFingerprint that can effectively and uniquely depict users. mFingerprint does not expose raw sensitive information from the mobile device, e.g., the exact location, WiFi access points, or apps installed, but computes privacy-preserving statistical features to model the user. These descriptive features obscure sensitive information, and thus can be shared, transmitted, and reused with fewer privacy concerns. By testing on 22 users’ mobile phone data collected over 2 months, we demonstrate the effectiveness of mFingerprint in user modeling and identification, with our proposed statistics achieving 81% accuracy across 22 users over 10-day intervals.