Client-side cheat detection in games using machine learning

Astract:

Statistics show that over the last years the popularity and the market share of online games has been increasing [Sta15c]. Playing with thousands of other players worldwide there is always the risk of people cheating to gain an advantage over others. This leads to an unbalance in a game and therefore unsatisfied customers from the game developers’ point of view. The goal of the gaming industry is to prevent that and to reveal as many cheats as possible to be able to consequently punish cheating players. There are many solutions for cheat detection, but the biggest problem with them is that they often rely on signatures to recognize cheats and therefore are unable to automatically adapt to new ones. Our goal is to reach a solution, in which the cheat detection program can learn to adapt to new situations without manual input. To fulfill these requirements machine learning presents itself as a suitable concept. Using Hidden Markov Models we will train the models to represent normal behavior and comparing test data to those models will yield whether a cheat has been used or not. Our assumption is that using cheats will result in deviations from the described normal behavior. With regular training the models can adapt themselves to code changes of a game and are able to even detect newly created cheats since they do not depend on prior knowledge of a cheat.