Perturbing Convolutional Feature Maps with Histogram of Oriented Gradients for Face Liveness Detection

Face anti-spoofing in unconstrained environment is one of the key issues in face biometric based authentication and security applications. To minimize the false alarms in face anti-spoofing tests, this paper proposes a novel approach to learn perturbed feature maps by perturbing the convolutional feature maps with Histogram of Oriented Gradients (HOG) features. The perturbed feature maps are learned simultaneously during training of Convolution Neural Network (CNN) for face anti-spoofing, in an end-to-end fashion. Extensive experiments are performed on state-of-the-art face anti-spoofing databases, like OULU-NPU, CASIA-FASD and Replay-Attack, in both intra-database and cross-database scenarios. Experimental results indicate that the proposed framework perform significantly better compare to previous state-of-the-art approaches in both intra-database and cross-database face anti-spoofing scenarios.