In our solution, there are client-side and server-side processes. Each side consists of multiple machine learning algorithms. Client-side processes are run with our client SDK which has been provided for iOS, Android and web. In the client SDK, we check that the face must be in the camera throughout the process and if we don't find the face in all the images, the process is stopped. We also check the user's gaze and head position in each frame.

The gaze pattern is used to check eye proximity (whether the user is asleep or not). Head pose verification is used to avoid data diversity. This makes our dynamic liveness model train better and easier. This detection of gaze and pose does not complicate the gesture of the user, on the other hand, it secures the system. In addition, in the whole procedure, a user must adapt his face to an oval. We generate the random oval in the random position of the screens. Displaying the random oval in random screen positions has two advantages. First of all, it prevents injection. Secondly, this movement makes some bold spoofing signals for attacks. In addition, we collect 30 client-side images.

On the server side, our dynamic face spoofing model, which is a convolutional model, was trained on our data. It uses 30 frames from the video and extracts a depth map capable of detecting depth and detecting spoofing attacks based on depth detection. The main challenge was providing data for training models. Our team collected a large set of data from our first attendance App. We also explored social media to collect a large set of video data, then we cleaned and created spoof data from it.

SDK are available for iOS, Android, Linux or Windows. If you want to use our face liveness detection software development kit, please send us a request.