It has become increasingly difficult to detect cheaters in mobile games. As the complexity of games grow, so do the different methods of cheating. As users realise the boundaries of simple mechanisms they find more nuanced ways to cheat the system. In this talk I will be taking you through a number of detection methods and discussing implications and implementation in the real world.
Outlier and anomaly detection is an interesting and much used area of statistics. Its application ranges from fighting fraudulent transactions in finance, detecting anomalies in data systems and in data ingestion quality assurance.
In this talk I’ll be describing an alternative example application of detecting cheaters within mobile games. I’ll be discussing the motivation around detection and what in particular we are trying to detect - detecting every facet of cheating behaviour is not necessary; cheating that gives a negative experience to other users is particularly undesirable.
Part of this talk will go into the various approaches to outlier detection: Single variate heuristics Density based approaches in multivariate data (e.g. minimum volume ellipsoid) Neural network methods (auto-encoders and replicator neural networks)
I will also explore the actual implementation issues and solutions in the particular example of mobile games. In the Free-to-play market, the majority of our users enjoy the game for free - with a minority making up the player base that spends on in-app purchases. Because of this, our highest value customers can often look like outliers alongside cheaters. Getting those sets of users mixed up is highly undesirable so some manual intervention is needed to make sure there are very little or no false positives.