Robust ECG based human verification technique

Abstract

Biometric plays a significant role in person verification. However, conventional biometric methods such as fingerprints and facial recognition are vulnerable to identity theft. Electrocardiogram (ECG) has emerged as a promising biometric modality due to its unique characteristics and inherent liveness detection criteria, making it difficult to forge. Preliminary studies have demonstrated the feasibility of ECG for biometrics recognition; however, challenges remain in ensuring its practicality, reliability, and usability in real-world scenarios. This thesis identifies three main research issues that are crucial for enhancing user acceptance of ECG based biometric recognition which are distinctiveness, collectability, and permanence. Existing studies predominantly focus on ECG biometrics in controlled environments. Therefore, this study explores biometric verification in two circumstances which are low-sampling ECG signals under different physiological conditions and compressed ECG signals without prior decompression considering four variables that are different activity settings, gender group, time variability and age categories. For low sampling data, Maximal Overlap Discrete Wavelet Transform (MODWT) is introduced as a novel signal pre-processing technique. The study managed to highlight the issue of distinctiveness where the outcome has demonstrated that different individuals under various physiological conditions possess each own QRS complex as unique identifiers with classification accuracy up to 99.7% by using Quadratic SVM. Secondly, to be able to perform biometric in a compressed state considering four different variables emphasises the following issues that are collectability and permanence. By using Symlet wavelet as the compression method, the classification accuracy by using Medium Gaussian SVM remains high. Different activity settings achieved a classification accuracy of up to 97.7%, demonstrating the robustness of ECG biometrics across various physical movements. Gender-based analysis shows an accuracy up to 98.6% where male participants show higher accuracy compared to female subjects. Time variability demonstrate high accuracy up to 98.8%. Additionally, age group analysis reveals that younger subjects achieve higher classification accuracy compared to older participants. The proposed approach demonstrates that even in a compressed state, ECG signals can still be effectively collected and measured and this addresses two important issues in biometric which are collectability and permanence. The study also discovered that, optimal compression of 50% can be achieved with high classification accuracy at level of decomposition of N=1. File size of the ECG signal also significantly decrease with each level of decomposition. Furthermore, each level of decomposition significantly reduces the ECG file size, making it more efficient for real-world applications without compromising verification performance. The robustness of the suggested technique is proven as the high classification results show that biometric recognition can be performed in compressed state. In conclusion, ECG based biometric is a promising approach to combat identity theft and complementing the current authentication system. By addressing the main issues of distinctiveness, collectability, and permanence, the proposed method suggests a robust and reliable mechanism for real life applications that offers a significant advancement in biometric recognition.