Additive Manufacturing of 3D Face Masks for Biometric Spoofing

L. Christie, A. Stone, K. Mancuso, D. Fedorishin, P. Schneider, K. Oh
University at Buffalo,
United States

Keywords: biometrics, additive manufacturing, spoofing


This research focuses on the use of additive manufacturing techniques, fused deposition modeling, stereolithography, and live casting, for the fabrication of synthetic face masks to challenge the biometric facial Liveness Detection scheme (Apple FaceID) seen on the iPhone X. Prior Work: Recently biometric security systems such as Apple’s FaceID and Qualcomm’s SENSEID have replaced passwords in mobile device security, leaving many people susceptible to biometric spoofing attacks [1]. In response, companies implement Liveness Detection schemes as counter attack measures. Apple’s approach utilizes the presence of autonomous ocular movement associated within the retina while the user is directing their attention to the device [2]. Using a volumetric regression networks, a single 2D face image can be morphed into a 3D model [3] from which a face “spoof”, or fake biometric, can be printed. Methods: Implementing a series of 3D additive manufacturing techniques, FDM, SLA, and live casting, a series of face masks were fabricated as biometric spoofs. Facial mask creation is done through a process flow that is outlined in Figure 1. The 3D face masks were created by implementing a volumetric regression network in order to predict the 3D features of the face from a single 2D image [3]. The reconstructed 3D print file is a low resolution mesh structure requiring mesh optimization implementation in order to smooth the face model for better accuracy. Three methods were used for creation of 3D synthetic facial masks and are shown in Figure 4. A predicted face model was created using the FolgerTech FT5. A facial reconstruction was done through and was printed via stereolithography printing. Experiments: 3D facial spoofs were created and tested using the iPhone X through 3D printing techniques. Quantification of systems performance was displayed in the form of a receiver operating curve (ROC) considering both false rejection and false acceptance rates. Evaluation will be performed using a multitask cascaded convolution neural networks described in [4] to detect the faces from a natural scene in addition to a squeeze and excitation enhanced neural network trained for facial recognition [5]. These algorithms can produce matching confidence scores to determine how similar the spoof is to the real face. The analysis of basic facial features can be seen in Figure 2. Low cost additive manufacturing has opened up new threat levels associated with facial biometrics. This research aims to quantify the effectiveness of 3D printed spoof masks in regards to defeating a biometric Liveness Detection System from something as easily accessible as a 2D picture taken from social media. Recognizer algorithms are applied to the faces to generate ROC curves and confidence scores. These algorithms are not robust to liveness detection and will serve as a baseline for spoof accuracy before moving into spoofs that can beat liveness detection such as blink and blood flow detection.