敌手分类的Stackelberg博弈分析

Stackelberg game based theoretical analysis for adversarial classification

在实际电子测量分类系统中,随着机器学习技术的广泛应用,它面临着来自对抗性环境中敌手攻击的巨大挑战。利用非合作二人Stackelberg博弈,并结合实际存在的跟随者反应时延问题,对敌手和数据挖掘者之间的交互行为进行了建模。和以前工作不同之处,在于本模型中考虑了跟随者反应时间对领导者和跟随者收益的影响。首先对Stackelberg博弈问题进行了形式化描述,然后对敌手分类问题进行了博弈建模,重点分析了此条件下博弈双方的收益,最后测试了本模型的正确性和有效性。实验结果表明,应该密切注意异常数据的变化,从而更好的对抗敌手攻击。

With the wide applications of machine learning technologies in practical classification systems of electronic measurement,they are facing the great challenges of adversarial attacks from adversarial environments.Combining with reaction-time delay of the follower,we model the interaction between the spammer and the data miner as a two-person sequential non-cooperative Stackelberg game in this paper.Different from the past researches,the influences of reaction-time delay for the leader and follower are considered.First we formulize Stackelberg game.Then we model adversarial classification problem as a Stackelberg game in which our emphasis is on analyzing the payoffs of the leader and follower under the follower＇s reaction-time delay.At last,the model＇s correctness and effectiveness are verified through our experiments.The experiment results show that we should pay more attention to the change of anomalous data so as to fight better against the adversarial attacks.