Human Authentication From Ankle Motion Data Using Convolutional Neural Networks

–  Published in:    2016 IEEE Statistical Signal Processing Workshop (SSP)

–  Authors:            Matteo Gadaleta, Luca Merelli, Michele Rossi

Paper Overview	Goal:  Use an ankle-worn inertial measurement unit (IMU) to authenticate users from their gait signatures. The authors present a framework integrates gyroscope information with accelerometer signal to get more accurate authentication result. Use Convolutional Neural Networks (CNN) as universal feature extractors, combination with a one-class classifier based on a Support Vector Machine (SVM). With a sequential decision maker that uses the scores outputted by the SVM across subsequent walking cycles.
System Model
Feature Extraction (Convolutional Neural Networks)	First convolutional layer: use linear activation functions to perform a preliminary (linear) filtering of the input signals. Second convolutional layer: introduce non-linearity through a tanh activation function. The max pooling layer: cuts down by half the number of output elements with respect to those in the second convolutional layer. Full-connected layer: output a feature vector. The last layer employs a softmax function..
Conclusion	Proposed an authentication framework that uses accelerometer and gyroscope data acquired from ankle-worn Inertial Measurement Units (IMU). It features a cascade of processing tools including, walking cycle segmentation, a convolutional neural network for feature extraction, a one-class classifier based on support vector machines and a sequential decision maker. Allowing the identification of the user in fewer than 5 steps and with false positive/negative rates smaller than 1%