Compared to traditional behavioral deviance, online deviant behavi or (like cyberbullying) is more likely to spread over online social communities since it is not restricted by time and space, and can occur more frequently and intensely. To control risks associated with the spread of deviant and anti-normative behavior, it is ess ential to understand online users' reaction when they interact with other users. In this paper, we model online users' behavior interaction as an evolutionary game on a graph and analyze users' behavior dynamics under different network conditions. Based on this theoretical framework, we then investigate behavior control strategies that aim to eliminate behavioral deviance. Finally, we use a real world dataset from a social network to verify the accuracy of our model's hypothesis. We also and test the performance of our beh avior control strategy through simulations based on both real and synthetically generated data. The experimental results demonstrate that our behavior control methods can effectively eliminate the impact of bullying behavior even when the proportion of bullying messages is higher than 60%.