As biometric recognition becomes increasingly popular, the fear of circumvention, obfuscation and replay attacks is a rising concern. Traditional biometric modalities such as the face, the fingerprint or the iris are vulnerable to such attacks, which defeats the purpose of biometric recognition, namely to employ physiological characteristics for secure identity recognition.
This thesis advocates the use the electrocardiogram (ECG) signal for human identity recognition. The ECG is a vital signal of the human body, and as such, it naturally provides liveness detection, robustness to attacks, universality and permanence. In addition, ECG inherently satisfies uniqueness requirements, because the morphology of the signal is highly dependent on the particular anatomical and geometrical characteristics of the myocardium in the heart.
However, the ECG is a continuous signal, and this presents a great challenge to biometric recognition. With this modality, instantaneous variability is expected even within recordings of the same individual due to a variety of factors, including recording noise, or physical and psychological activity. While the noise and heart rate variations due to physical exercise can be addressed with appropriate feature extraction, the effects of emotional activity on the ECG signal are more obscure.
This thesis deals with this problem from an affective computing point of view. First, the psychological conditions that affect the ECG and endanger biometric accuracy are identified. Experimental setups that are targeted to provoke active and passive arousal as well as positive and negative valence are presented. The empirical mode decomposition (EMD) is used as the basis for the detection of emotional patterns, after adapting the algorithm to the particular needs of the ECG signal. Instantaneous frequency and oscillation features are used for state classification in various clustering setups. The result of this analysis is the designation of psychological states which affect the ECG signal to an extent that biometric matching may not be feasible. An updating methodology is proposed to address this problem, wherein the signal is monitored for instantaneous changes that require the design of a new template.
Furthermore, this thesis presents the enhanced Autocorrelation- Linear Discriminant Analysis (AC/LDA) algorithm for feature extraction, which incorporates a signal quality assessment module based on the periodicity transform. Three deployment scenarios are considered namely a) small-scale recognition systems, b) large-scale recognition systems and c) recognition in distributed systems. The enhanced AC/LDA algorithm is adapted to each setting, and the advantages and disadvantages of each scenario are discussed.
Overall, this thesis attempts to provide the necessary algorithmic and practical framework for the real-life deployment of the ECG signal in biometric recognition.
Identifer | oai:union.ndltd.org:TORONTO/oai:tspace.library.utoronto.ca:1807/31673 |
Date | 05 January 2012 |
Creators | Agrafioti, Foteini |
Contributors | Hatzinakos, Dimitrios |
Source Sets | University of Toronto |
Language | en_ca |
Detected Language | English |
Type | Thesis |
Page generated in 0.002 seconds