Adversarial Gait Detection on Mobile Devices Using Recurrent Neural Networks

摘要

This paper presents an implicit and continuous user verification service, called dCASTRA, for mobile devices based on walking patterns inferred from smart phone sensors. We use LSTM (Long Short Term Memory) neural networks for learning gait biometrics from raw accelerometer and gyroscope data and enable a device centric implementation of the deep learning models for faster predictions. One of the challenges in building a gait biometric model is to differentiate the sensor data pertaining to the walking activity from other activities such as sitting, standing, climbing, running and driving, etc. We design a multi-layer framework, where the initial layer relies on Google Activity Recognition Service to extract the segments corresponding to the walking activity with high confidence and feed extracted time series data to LSTM networks in the subsequent layer. The use of LSTMs eliminate the need for tedious feature engineering and further enable us to capture long-term dependencies within temporal sequences, often overlooked by existing efforts. We use Google TensorFlow to develop LSTM based gait biometrics and deploy on Android-based smart phones for real-time prediction and evaluation. We compare dCASTRA with our prior effort and with other deep network architectures such as Convolutional Neural Networks (CNNs). Our results manifest that LSTM and CNN based dCASTRA identifies users in an average 5-6 steps (using 50 Hz sensor sampling rate) with 99% detection accuracy. However, CNNs face significant training overhead as opposed to LSTMs which in turn limits its ability to be deployed in practice.