Improved Face Spoofing Detection Using Random Forest Classifier with Fusion of Luminance Chroma Features

Abstract

Ambient computing applications verify identity of an individual using biometric identity in addition to conventional security measures. The cyber applications do have provision of face identification to strengthen the secure access for financial and other critical applications. Face recognition is more susceptible to spoofing / face presentation attacks, where the print of face or face video replay is used to spoof the identity of an individual. The paper proposes performance improvisation of existing face presentation attack detection technique using machine learning algorithms and fusion of luminance-chromaticity (Kekre-LUV, CIE-LUV, YCrCb) face features. The paper does empirical assessment of color space combinations that are used for feature extraction to decide whether face is real or spoofed. Along with the earlier advocated use of ExtraTree classifier, the paper explores using Random Forest and other ensembles of machine learning algorithms (classifiers) in detection of face presentation attack and Random Forest significantly improves the performance of face spoofing detection as it is clearly evident form articulated results. The paper also proposes use of Kekre-LUV color space which is computationally lighter than earlier used CIE-LUV, experimental analysis shows that almost similar performance of face presentation attack detection is observed using Kekre’s LUV color space. Further the fusion of the luminance chroma features are proposed to be used for higher accuracy. The proposed method is validated using two datasets as ‘Replay Attack’ and ‘NUAA’, with help of ‘accuracy’ and ‘half total error rate’ (HTER) as performance measures. The achieved accuracy and HTER have proved the worth of proposed methodology.