Trajectory-Based Mobile Game Bots Detection with Gaussian Mixture Model.

Recent developments in mobile games have heightened the need for mobile game bots detection, while related researches are scarce. Compared with PC games, detecting game bots in mobile games is more challenging in the following aspects: (I) Privacy. User information in mobile games is more sensitive than on PC. It is unacceptable to employ the program logs or locations to detect the bots. (II) User operations in mobile games are complex. To achieve the same goal, the operations may be completed by multiple fingers and can be various in different users. (III) The labeled data samples are few. The bots in mobile games change frequently, leading that the labeled samples for recent bots could be extremely rare. Unfortunately, these problems have not been well-solved in recent literatures. Thus, in this paper, we propose a reconstructionbased model, namely Mutil-view GMTVAE, which utilizes the finger touch records collected by screen sensors to infer the potential cheating players in a semi-supervised way. Mutil-view GMTVAE models the complex operation records in a latent space with a VAE-enhanced GMM which learns more general representations for records reconstruction. Extensive experiments on two NetEase games show that Mutilview GMTVAE achieves better performance than the baselines and is general to detect bots in different mobile games.