A Robust Alternative for Spoof Detection Using GMM

• Activity Number: 137
• Type: Contributed
• Date/Time: Monday, August 1, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Defense and National Security
• Author(s): Umashanger Thayasivam*+ and Ravi P. Ramachandran and Sachin Shetty
• Companies: Rowan University and Rowan University and Tennessee State University
• Keywords: spoof detection ; GMM ; biometric ; robust ; SVM ; HMM

Abstract

Biometric technologies have an essential role in assuring and safeguarding personal, national and global security. Such is the value of the assets or information that they protect, biometric systems present a serious and growing target for criminal attack. One form of attack involves so-called 'spoofing' where a person attempts to masquerade as another by falsifying data in order to gain an illegitimate advantage. Alarmingly, as widely acknowledged in the literature, the threat to biometric technologies from spoofing attacks is all too real. A spoof detection system (SDS) is imperative to counter the possibility of hackers using record-and-play spoof attacks to compromise the biometric speaker recognition system. To implement the SDS, we will first evaluate the effectiveness of the popular stochastic based anomaly detection models namely, Hidden Markov Models (HMM), Gaussian Mixture Models (GMM), and Support Vector Machines (SVM). We also propose a robust alternative with L2E distance approach based on GMM. We will perform the experiments on the same testbed and choose the appropriate model based on the following low false positives, quick spoof detection and response time.