Feature Selection, Cross-view Matching and Spoof Detection for Electrocardiogram-based Human Recognition

Biometric technology for identifying individuals has been extensively used all over the world. As this technology becomes increasingly popular, the potential of fooling or spoofing it is a rising concern. For example, fingerprint may be left behind whenever we touch a glass surface like a smart phone and from that some synthetic samples can be created that can be used to fool the biometric system. Iris images can also be captured from a few meters distance, not to mention face images that can be captured from a longer distance. Nowadays biometric spoofing and countermeasures is an active research area and there has been a lot of efforts towards a promising approach to ensure the presence of a real legitimate user. Medical biometrics such as Electrocardiogram (ECG) have interesting advantages over conventional biometrics. ECG is a vital signal and it naturally provides liveness detection. In addition, compare to conventional biometrics, if not impossible it is far more difficult to steal someone's ECG. Since ECG can be recorded from fingertips it is a natural choice for fusion with fingerprint. In this research work, we first provide a practical framework for fusion of ECG and fingerprint such that the resulting multimodal biometric system is more robust against novel spoof attacks. Despite its advantages, ECG comes with its challenges. ECG is a non-stationary signal and factors like emotion, diet and hear rate affect its waveform especially when recorded in different sessions. It becomes more challenging when subjects are enrolled using only one session and tested on another session that could be a few weeks apart. To reduce the effect of such non-stationarity, we propose a feature selection method to select a subset of features that are more stable across different recording sessions. Furthermore, to cope with the effect of heart rate on performance of ECG recognition, we propose a novel method for cross-view matching where the distance between two samples situated in different “views” can be computed. In this context “view” refers to different heart rates where samples in different views can be represented by feature sets that can be different in type and length.Ph.D.2019-12-19 00:00:0