Disguised face liveness detection: an ensemble approach using deep features

Abstract

Deep learning models have surpassed classic machine learning models in face anti-spoofing detection during the last decade. Most face-spoofing detection algorithms are biased toward a single presentation attack, failing to robustly detect multiple spoofing scenarios, especially when faces are disguised in surveillance. Face anti-spoofing techniques have attracted a lot of interest since they try to determine whether the captured face is real or spoofed. We propose an ensemble technique that uses deep features to determine whether disguised faces in surveillance videos are live or spoofed. Integrating the precisely specified features and automatically extracting features using deep learning, helps to improve the performance of the model, reduce errors, and avoid overfitting. Local Binary Pattern (LBP) images are generated by using Handcrafted features extracted using the Local Binary Pattern (LBP) descriptors. Deep features are extracted using SEResNeXT50 and EfficientNet-B3 pre-trained Convolution Neural Network (CNN) architecture from raw and LBP images, respectively, and are fused for identifying the disguised face liveness in surveillance videos. The effectiveness and resilience of the proposed method are proved by experimental results on numerous benchmark datasets and achieved an ACER of 0% on Replay-Attack and MSU-MFSD datasets.

Keywords

• Deep learning features
• disguised faces
• ensemble
• face anti-spoofing
• handcrafted features
• local binary pattern