Publication

Abstract : The usage of wearable devices and smartphones is surging day by day, while most of them are equipped with Inertial Measurement Units (IMUs). IMUs can collect users’ private data and store and share it among different devices. The private data stored in mobile phones raises security concerns, which can be dealt with by user authentication mechanisms such as bio-metric. Therefore, a corrective authentication mechanism is required to ensure private data security. Gait analysis is one of the user authentication schemes. A gait authentication can be implemented using video and motion sensor data. Video data-based approaches are vulnerable because of the ease of imitating the videos, whereas the motion sensor data provide a more secure mechanism against attacks. This paper proposes a lightweight user gait authentication mechanism using motion sensor data collected using IMUs. Statistical features were extracted from the data of 15 users to model the authentication system and further employed a feature selection method to select the best features describing the data to avoid a significant correlation between the features. Separate models were developed for each user by following the one versus rest approach using the features mentioned above. In the first experiment, the modelling problem was formulated as a one-class classification problem due to the high imbalance between each class. In the third experiment, for modelling each user’s authentication system, an equal number of data points were collected randomly from the remaining users to prepare a balanced dataset. Both second and third experiments were modelled as a binary classification problem. Results showed that the Support Vector Machine (SVM) obtained the best authentication performance in the third experiment. The mean Equal Error Rate (EER) of all the users was also minimum for the SVM model trained using the balanced data in the third experiment.