New research Computer scientists from California University in Berkeley, Rwth Aachen in Germany and Unanimous Ai shows how people in virtual reality can be identified by body movement.
Scientists have collected telemetry data from more than 55,000 people who played in Beat Saber, the VR rhythm-game, in which the players wave with hand controllers to the music. Then they processed 3.96 TB of data from the BEATLEADER leaders table of 2 669 886 game repetitions from 55,541 users during 713 013 separate game sessions.
These Beat Saber Open Replay (BSOR) files contained the following information:
- metadata (devices and gaming settings);
- telemetry (measurements of the position and orientation of the hands, head, etc.);
- contextual information (type, place and entry time to the game);
- performance statistics.
The data obtained from the movements of the head and hands of the Beat Saber players allowed scientists to train the classification model in 5 minutes. This model, given the 100 minutes of data about the movement from the game, was able to unequivocally identify the player in 94% of cases. And in just 10 seconds of calculation of data on movement, the accuracy of the classification model was 73%.
The study shows that more than 55,000 “anonymous” VR games can be deanonymized to a specific person by simple tracking of their head and hands within a few seconds.
Experts concluded that the models of movement are so unique to humans that they can serve as an identifying biometry along with the recognition of faces or fingerprints. This really changes our idea of confidentiality in the metavselnaya, because, simply moving in virtual reality, you constantly broadcast your face or fingerprints.
To preserve the privacy of the user, scientists proposed introducing VR projects that change the performance of the players. Such projects should distort the movements of the players in order to hide the identified properties without preventing the process of players in the game.