In general, the accuracy of automatic recognition consists of two parts:
1. The accuracy of the event recognition algorithm, for example, the absence of a student's face on the screen, the presence of another person in the room, unknown voices, switching windows or tabs on the desktop, etc.;
2. The accuracy of interpreting an event as an attempt to cheat.
More complex and resource-intensive models can improve the accuracy of photo/video recognition. For example, face recognition services used in banks cost at least $ 0.5 per transaction. During one exam, Examus algorithms use facial recognition more than 100 times. Therefore, the accuracy improvement of our algorithms is limited for economic reasons. Currently, the accuracy of the neural network we use to recognize faces is ~0.9947%.
Nevertheless, even if the algorithm's accuracy is 100%, there are still many questions about the interpretation of the events. For example, an automatically found "no face" event can be triggered by such cases as:
• A face is covered with hands
• A person turns away from the camera
• A person partially moves out of the camera's view – for example, only the left side of the face is visible.
What should be done in this case? Shall we generate the "Student is absent" event or not? On the one hand, the student is sitting still, and the proctor can see this. On the other hand, it can be suspicious behavior. If we decide to exclude such situations from the list of suspicious events, we have to use more complex neural networks, which is too costly for the reason described above.
Thus, when the AI proctor sees a suspicious event and marks it, a human proctor has to check this event to determine whether it is cheating or not.
Sometimes, by watching a video, a live proctor can determine the fact of cheating. In this case, there is a theoretical possibility of training automatic algorithms to identify such situations. The only question is complexity.
However, when there are no obvious actions in the frame, even a person cannot always confirm if a student is cheating. Therefore, it is almost impossible to teach a cyberproctor to do this.