MISIS University created a neural network to verify the authenticity of facial images. In NUST MISIS you can upload images for verification through a specially developed web application, where information about this is provided to socialbites.ca; There is also the possibility of real-time analysis via a computer camera.
In their work, the developers relied on presentation attacks: photographs of faces in printed and electronic form, as well as three-dimensional masks that imitate human facial features.
They analyzed five existing neural networks, selected the two most promising, and developed a version, a two-stage system, based on their experiments and observations.
“An important step in developing a machine learning solution is finding a dataset to train the models on. We used a dataset of 16,500 images: real and fake, with a roughly even distribution of facial recognition system deception types: photos printed on the screens of electronic devices, masks, and cartoon characters, and Using images. Alisa Semenova, one of the developers, told socialbites.ca: They also printed photos of people with various external features, prepared “fake” images and added them to the selection.
The first stage of face recognition uses a pre-trained MTCNN neural network that determines the position of the face in the image. A special area is then added to the image, 60% of whose area is the analyzed face. This approach provides a significant increase in accuracy. The InceptionResnet network is then used to provide numerical representations of facial features. In the second stage, several more layers of the neural network are used to select image features.
The results of the two stages are combined and passed through several final layers to reach a final conclusion about the authenticity of the image. This approach allowed the team to achieve a high level of accuracy in determining the authenticity of faces.
Previous researchers taught Neural network for interviewing sales managers.