Bot-pelganger: Predict and Preserve Game Bots' Behavior

Abstract

In most multiplayer online games, players' repetitive tasks (i.e., spec-up) are required to grow their characters. However, some users use illegal programs, “game bots,” to achieve a high level fast or gain cyber-money. Various methods have been proposed to identify game bots. However, the methods have generalization issues. Because the methods use features only existed in the specific game. Thus, we carefully use common features that existed in multiple datasets broadly, such as ‘login’ or ‘exit’ events to detect bots. Choosing such general events gives merits from the applicability view; however, if we only use time or space-related features, we fail to detect bots from normal users because the bots' behavior patterns are omitted too much. We use a convolutional LSTM (ConvLSTM) model to overcome this problem, superimpose their behavioral histories over time, and record them as image sequences. By finding a user who shows high self-similar behavior, we regard it as an unidentified bot; then, we update their behavior patterns for future use. As a result, the proposed model showed a high accuracy of 98% in classifyina game bot users.