Fingerprinting Smartphone Accelerometers with Traditional Classifiers and Deep Learning Networks

Abstract

Fingerprinting smartphones using their accelerometers has several applications, including activity recognition, driving style classification and device to device authentication. In this work, we study accelerometer-based smartphone fingerprinting. We gather data from mobile devices placed together to record identical vibrations. Then, we extract time domain features, which we use to train multiple traditional machine learning algorithms based on statistical properties of the data. Finally, we use the raw data in a more complex Convolutional Neural Network and compare the results. To make the investigations more challenging, we discuss fingerprinting both distinct and identical smartphones and reach an accuracy close to 100% with several traditional classifiers.