Altered Fingerprint Detection and Reconstruction using Artificial Neural Networks

Abstract

The constant use of fingerprints as a biometric modality has become even more widespread in recent years in terms of its application for person identification purposes. The integrity of the biometric fingerprint image database for identification purposes is also of paramount importance. Fingerprint alteration is a challenge that poses enormous security risks. Researchers have attempted to tackle these security risks, particularly for fingerprint alteration detection, reconstruction of the altered fingerprints, and gender classification, but with limited success so far, due to the lack of available altered fingerprint datasets. This research has collected, organised and made publicly available the Sokoto Coventry Fingerprint (SOCOFing) and Sokoto Coventry Fingerprint Dermalog Datasets (SOCOFing_Derma), which can now be used by researchers to test novel algorithms and systems for fingerprint and hand biometric systems. The research first investigated possible forensic techniques for the stored biometric fingerprint templates tampering detection using a Support Vector Machine (SVM) classification approach. Original and fake fingerprint images are used to train the SVM classifier. The fingerprint datasets from the Biometrics Ideal Test (BIT), which contains both fake and original fingerprints, are used for training and testing the classifier. The proposed approach detects alterations with high accuracy (94%). A case study was undertaken to assess problems in a real-life fingerprint biometric system, and this yielded the new fingerprint datasets referred to above, which are intended for use in the development of improved systems. Following this, deep Convolutional Neural Network (CNN) architectures are adapted for fingerprints alteration detection, gender classification, and the identification of hands and individual fingers. Transfer learning is employed to speed up the training of CNNs. Using our newly collected and developed datasets, the proposed CNN model for the detection of fingerprint alterations achieves an overall classification accuracy rate of 98.55%, in comparison with a ResNet18 model pre-trained on ImageNet, refined and later tested on the dataset, which achieved an accuracy rate of 99.86%. Most significantly, there is no known previous studythat has been able to effectively address these and other types of alterations, as achieved in this study. Gender classification is paramount for reducing the time that it takes to investigate criminal offenders and gender impersonation. The proposed CNN approach classifies andachieves an accuracy of 75.2%, 93.5%, and 76.72% for gender, hand, and fingers classification, respectively. Finally, an effective Deep Stacked Convolutional Autoencoder (SCAE) was proposed to restore the altered fingerprints to the original fingerprints. These results were obtained by using our Sokoto Coventry Fingerprint (SOCOFing), and Sokoto Coventry Fingerprint Dermalog Datasets (SOCOFing_Derma), developed and made publicly available through this research, and which may forthwith serve as a benchmark for fingerprint classification results.

Date of Award	May 2020
Original language	English
Awarding Institution	Coventry University
Supervisor	Vasile Palade (Supervisor), Michael Odetayo (Supervisor), Anne James (Supervisor) & Rochelle Sassman (Supervisor)