A Concept Drift Adaption Framework for Online Game Chargeback Detection

Abstract

The burgeoning online gaming market, driven by the rapid advancement of internet infrastructure and hardware performance, has increasingly become a focal point for criminal activity. Service providers in this industry are particularly vulnerable, suffering significant financial losses due to malicious chargebacks. Current countermeasures remain largely reactive, as providers typically block compromised game accounts only after an attack has occurred. Although prior studies have attempted to leverage machine learning techniques to detect fraudulent chargebacks, perpetrators often employ evasion strategies to bypass detection, thereby exacerbating the problem of concept drift in game records. To address this issue, we propose enhancing the detection of malicious behavior through behavioral analysis in online gaming. This study applies machine learning models to detect malicious chargebacks in online games, focusing not only on improving detection capabilities but also on introducing mechanisms for identifying and mitigating concept drift. We propose an adaptive learning model specifically designed to handle concept drift in the context of top-up fraud in online gaming, aiming to proactively prevent malicious chargebacks before financial losses occur. The final experimental results demonstrate that, by employing incremental learning methods after detecting concept drift, the LSTM-Seq2Seq model with an attention mechanism achieved the best MCC performance. Through the proposed incremental learning model, online gaming service providers can effectively enhance their detection capabilities and further reduce operational losses.