Face anti-spoofing is an AI-based technique to prevent false facial identification using a video, mask, or picture of the authorized individual. For instance, biometrics is the most common scenario, where the attacker uses an authorized person's fingerprints. The same can happen for facial recognition too. Spoofing is the most widespread technique to attack the facial recognition system, and hence, anti-spoofing should be prioritized to build robust solutions. We created one such secure and powerful approach to prevent spoofing attacks that have become prevalent these days.
Our client owns an IT and eCommerce-based business and provides cloud services to execute monetary transactions. Moreover, he offers alerts and payout solutions for e-payment platforms, lottery players, custom solutions, and e-subscription platforms. Our client was seeking professionals in the facial recognition system. The client had concerns regarding the ongoing spoofing attacks, and to secure his system, he approached HData Systems to build a robust solution against such spoofing attacks.
We began with model prototyping and shifted to analyzing commercial implementation. We divided the model training process into 2 phases - detecting video attacks & static image attacks with digital pictures.
We determined a genuine or fake person for static image attacks by using the frame sequence or one frame from the current dataset. Our experts instructed the model to categorize facial keypoints sequences as fake or real to be used later for the interface.
Our team determined a genuine person or fake one with the single frame or video frame sequence for digital attacks. We instructed a separate model to scan each frame and blend different features with the frame scanning output inside the final network using deep learning.
Lastly, we recommended a frame-by-frame model to evaluate each frame separately and draw specific features such as an RNN to examine the key points' dynamics and features.
Our team struggled with coaching the deep learning model for anti-spoofing, as our client required a model to detect the below attacks-
Digital attack
Image attack
Replay attack
Hence, our HData team has to create the data processes collecting particularly for that task and gather the data needed for training.
After our team successfully implemented the facial anti-spoofing solution for the client, it delivered terrific results. Our expert engineers provided the entire model training pipeline and interface, with the option to import the final neural network for the client to use from Java.
Our client received great suggestions from our professionals for advanced model training as our system addresses exact business needs concerning security.