Human-computer interaction on mobile devices driven by statistical models

OData support
Supervisor:
Dr. Gyires-Tóth Bálint Pál
Department of Telecommunications and Media Informatics

Nowadays portable devices - like smartphones, smartwatches - have become an important part of our daily lives. These devices contain an increasing amount of personal information, and they can even replace our credit cards. In order to protect these information, we have to take certain security precautions. Currently the sensitive data in mobile phones is most often protected with PIN codes. This method requires the user's explicit input, and the users have to memorize their own passwords. With the increasing number of used devices, the number of passwords to be memorized is also growing. To resolve this problem, alternative security approaches are being studied, one of these methods is biometric user authentication. Biometric authentication methods have many attractive features, such as user-friendliness, no memorization and no risk of identity-theft. I have chosen the topic of my research to be gait recognition. In my thesis I have reviewed the scientific literature connected to my topic of research.

I created a Google Android based data-recording application, which I have used to record gait samples from 14 persons. I preprocessed the recorded data with Fast Fourier Transformation, autocorrelation method and by examining the power of the signal’s frequency bands. I used the preprocessed data to train the Random Forest, SVM, LDA and KNN machine learning algorithms. I created an experimental system, in which I implemented different combinations of the learning algorithms and the preprocessed training data. I used the cross-validation method for the evaluation of the experimental system. According to the tests, the system can recognise a user with 94% accuracy, and it falsely identified users in 1,09% of cases.

Downloads

Please sign in to download the files of this thesis.