Please use this identifier to cite or link to this item:
|Title:||Development of Biometric Verification Algorithm using Electroencephalogram (EEG)|
|Supervisor:||Singla, Sunil K.|
|Abstract:||The biometrics is one of the most innovative areas of Science dealing with the automatic identification/verification of the person in the modern world. It involves the measurement of physiological and behavioral characteristics of a person for authentication. The conventional biometrics (e.g. fingerprint, voice, hand geometry etc.) are prone to forgery and can be spoofed by imprinting artificial finger, voice imitation, and fake signature etc. Moreover, the presence of an object is not guaranteed in conventional biometric systems. Such problems lead to uncover alternative and effective biometric traits which are more accurate, reliable and secure for authentication of an individual. The bio-signals are unique, confidential, secure, almost impossible to mimic, hard to be copied and guarantees the individuals presence during acquisition of signals. In this thesis, EEG signals were used for biometric verification because, apart from basic characteristics of a biometric system (i.e. universality, uniqueness etc.), it possess attributes of liveliness detection due to presence of a legitimate individual for signals acquisition. Moreover, the EEG records brain waves which are unique and cannot be read by others. In this way, EEG is non-vulnerable to spoof attacks and thus, highly reliable for person identification/verification. Keeping in mind the specific traits of EEG signals, this thesis was focused on the development of a system for person verification using EEG signals. The entire study is briefly explained below. The database consists of 1960 samples (acquired/standard) from 67 subjects in three categories: (i) Relaxed State with Eyes Open: 32 human volunteers with 30 trials per person; (ii) Alcoholic/Controlled Disposition: 30 subjects with 30 trials (Standard: Ingber); (iii) Mental Ability Tasks: 05 subjects performing 05 different cognitive tasks (Standard: Keirn and Aunon). After the acquisition, the EEG signals were pre-processed for artefacts elimination (caused by eye blinks) using Fast Independent Component Analysis (FastICA), a multivariate signal statistical technique. It (FastICA) finds the underlying components from the mixture of signals (which are statistically independent and non-gaussian) and decomposes the signals into several components. The application of FastICA resulted in refined EEG signals with an improved Signal to Noise Ratio (SNR). After pre-processing, 20 features were extracted using Linear (Time and Frequency domain) and Non-Linear (Fractal Dimension) techniques. Five features (e.g. RMS, Approximate Entropy, LZ Complexity, Median Power Frequency and Spectral Edge Frequency) were selected amongst 18 features under linear techniques based upon minimum intra-individual and maximum inter-individual differences using one-way Analysis of Variance (ANOVA). Non-linear techniques (Higuchi Fractal Dimension and Correlation Dimension) were applied on third type of dataset (i.e. mental ability tasks). For verification, the individual’s query samples were compared to previously enrolled data of that person. The verification was carried out on individual’s extracted features of time and frequency domain at three levels of security threshold i.e. 5%, 10% and 15% (Higher to Lower security). The maximum Genuine Acceptance Rate (GAR) of 86.12% (LZ Complexity) was achieved in the case of relaxed subjects, whereas, GAR of 87% (SEF) in Alcoholic/controlled disposition. After normalization of features using Minimum Maximum and Euclidean Norm techniques, a further improvement was observed. The normalized datasets were assigned with weights (estimated from Rank Order Centroid method) to find the cumulative score, resulting in maximum GAR of 91.04% (with Euclidean Norm) in relaxed subjects whereas, 91.40% (with Min-Max scaling) in Alcoholic/controlled disposition. Also, an advance machine learning technique (Support Vector Machine-SVM) resulted in Correct Classification Rate (CCR) of 97.51% (FRR-2.49%) and 96.82% (FRR-3.17%) in relaxed and alcoholic/controlled subjects respectively. Under non linear technique using SVM, the CCR of 97% was observed with HFD features; whereas, 95% with Correlation Dimension in the case of mental ability tasks. To overcome the limitations of single EEG biometric such as universality (EEG not recorded for epileptic or Alzheimer’s patient), uniqueness etc., the EEG data were combined with the Fingerprint at score level using the Fuzzy logic technique. The fusion of the two biometric makes the system more robust, flexible, secure and accurate. The overall Genuine Acceptance Rate (GAR) of 93.2% and FAR of 7% was achieved with the fusion of EEG and Fingerprint.|
|Appears in Collections:||Doctoral Theses@EIED|
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.