Deep learning and multivariate time series for cheat detection in video games

Abstract

Online video games drive a multi-billion dollar industry dedicated to maintaining a competitive and enjoyable experience for players. Traditional cheat detection systems struggle when facing new exploits or sophisticated fraudsters. More advanced solutions based on machine learning are more adaptive but rely heavily on in-game data, which means that each game has to develop its own cheat detection system. In this work, we propose a novel approach to cheat detection that doesn't require in-game data. Firstly, we treat the multimodal interactions between the player and the platform as multivariate time series. We then use convolutional neural networks to classify these time series as corresponding to legitimate or fraudulent gameplay. Our models achieve an average accuracy of respectively 99.2% and 98.9% in triggerbot and aimbot (two widespread cheats), in an experiment to validate the system's ability to detect cheating in players never seen before. Because this approach is based solely on player behavior, it can be applied to any game or input method, and even various tasks related to modeling human activity.