Spoofing Real-world Face Authentication Systems through Optical Synthesis

Facial information has been used for authentication purposes. Recent face authentication systems leverage multi-modal cameras to defeat spoofing attacks. Multimodal cameras are able to simultaneously observe the targeting people from multiple physical aspects, such as the visible, infrared, and depth domains. Known spoofing attacks are not effective in evading the detection since they cannot simulate multiple modalities at the same time.This paper presents a new class of spoofing attacks on multimodal face authentication systems. Its main idea is to forge each and every modality and then combine them together to present to the camera. The attack is realized with a special display device called Hua-pi display. It costs less than $500 and incorporates dedicated scene generators to optically reproduce multimodal scenes of an authorized user, and then synthesizes the scenes together at the camera’s view point through optical combiners to fool face authentication systems. We evaluate the risks of this attack by systematically testing it against the latest commercial face authentication products from major vendors in the field. The results not only demonstrate a successful bypass rate of 80% but also characterize the impacting factors and their feasible regions, revealing a new and realistic threat in the field.