A new face presentation attack detection method based on face-weighted multi-color multi-level texture features

Biometric data (facial, voice, fingerprint, and retinal scans, for example) are widely used in identification due to their unique and irreversible nature. Facial recognition technologies are employed in a wide range of applications due to their contactless nature and convenience. However, technological advancements and the availability of access to personal information have rendered these biometric systems susceptible to attacks utilizing fake faces. As a result, the issue of anti-spoofing has emerged as a critical one in the field of facial recognition. This study proposes a joint face presentation attack (FPA) detection method based on face-weighted multi-color multi-level LBP features extracted from the combination of device-dependent HSV and device-independent L*a*b* color spaces. The facial images were converted to HSV and L*a*b* color spaces. Three levels of regional LBP features were extracted from each color channel and then concatenated. Finally, a Multi-Color Multi-Level LBP (MCML_LBP) feature vector was obtained. In addition, the Face Weighted MCML_LBP feature vector was produced (FW_MCML_LBP) by adding the LBP histogram extracted from the central region of the normalized image. The feature vectors are used to train an SVM classifier after reducing their size using PCA. Twenty-five different test scenarios were subjected to experimentation on the CASIA and Replay-Attack databases. 2.11% EER and 0.19% HTER were achieved on CASIA (Overall) and Replay-Attack (Grandtest) databases, respectively, using the L*a*b color space and the proposed feature extraction method. The results of the study showed that the proposed method was successful in FPA detection compared to the state-of-the-art methods.