Face Presentation Attack Detection

Presentation attacks, also known as spoofing attacks, are deliberate attempts to deceive biometric systems by presenting fake inputs that bypass security mechanisms. Among these, face presentation attacks involve using fraudulent face inputs such as photos, videos, or 3D masks of legitimate users to deceive the camera sensor. This paper focuses on detecting such attacks by presenting a novel approach built on supervised learning techniques. The YOLO model was trained on 7,000 images in total, enabling it to distinguish between live and spoofed inputs with an accuracy of 94.5% when tested. The key insight leveraged by the model is the difference in gradient smoothness: the natural tone transitions in live faces are significantly different from those observed in spoofed images or videos. This technique not only enhances the detection capabilities of biometric systems but also holds potential for applications in securing online transactions by adding a layer of liveness validation during authentication processes.