Date of Award

5-2021

Document Type

Thesis

Degree Name

Master of Science (MS)

College/School

College of Science and Mathematics

Department/Program

Computer Science

Thesis Sponsor/Dissertation Chair/Project Chair

Michelle Zhu

Committee Member

Tianyang Wang

Committee Member

Weitian Wang

Committee Member

Jiayin Wang

Abstract

Fingerprints play a significant role in many sectors. Nowadays, fingerprints are used for identification purposes in criminal investigations. They are also used as an authentication method since they are considered more secure than passwords. Fingerprint sensors are already widely deployed in many devices, including mobile phones and smart locks. Criminals try to compromise biometric fingerprint systems by purposely altering their fingerprints or entering fake ones. Therefore, it is critical to design and develop a highly accurate fingerprint classification. However, some fingerprint datasets are small and not sufficient to train a neural network. Thus, transfer learning is utilized. A large Sokoto Coventry Fingerprint Dataset (SOCOFing), which contains 55,273 fingerprint images, was first used to train a convolutional neural network model to detect image alteration and level of alternations. The model was able to achieve an 81% of accuracy. Then, a few layers of SOCOFing model were used and adapted to train another smaller dataset, namely ATVS-FakeFingerprint Database (ATVS-FFp DB), which contains 3,168 fingerprint images. Two models were trained. The first transferring model was built to classify images into real and fake, and a remarkable classification accuracy of 99.4% was achieved. The second transferring model was used to detect if the image was fake and if the user was cooperating in the generated faked fingerprint. The model achieved a classification accuracy of 97.5%. The transfer learning technique proves to be very effective in addressing insufficient dataset issues for deep learning.

File Format

PDF

Download

Share

COinS