摘要
针对数据库中知识发现的对抗性分类博弈进行研究。将该问题构建为一个代价敏感的分类器和代价敏感的对手之间的博弈;基于朴素贝叶斯,分别提出以使预期效用最大化的对手对抗分类器的最优策略和分类器对抗对手策略的最优策略,提出基于剪枝规则和贪婪向前搜索的高效算法计算或近似这些最优策略,实现分类器适应对手不断变化的策略。仿真结果表明,所提分类器相比标准的NB分类器在效用收益、假阳性和假阴性分类性能方面分别提高约47.8%、24.1%和29%,与对手共同进化的能力更佳。
Research on the adversarial classification game of knowledge discovery in database was carried out.The problem was formulated as a game between a cost-sensitive classifier and a cost-sensitive adversary.Based on Naive Bayes(NB),the optimal strategy of adversary against classifier and the optimal strategy of classifier against opponent strategy to maximize their expected utilities were proposed,respectively.Efficient algorithms based on pruning rules and greedy forward search were proposed to calculate or approximate these optimal strategies.The classifier thus could adapt to the adversary’s constantly changing strategies.Simulation results show that,compared with the standard NB classifier,the proposed adversary-aware classifier improves the utility gain,false positive and false negative classification performance by about 47.8%,24.1%and 29%,respectively,as well as the ability to co-evolve with the adversary.
作者
郝惠惠
王林
HAO Hui-hui;WANG Lin(College of Information Engineering,Zhengzhou Technology and Business University,Zhengzhou 451400,China;School of Automation and Information Engineering,Xi’an University of Technology,Xi’an 710048,China)
出处
《计算机工程与设计》
北大核心
2023年第4期1136-1143,共8页
Computer Engineering and Design
基金
河南省高等学校重点科研基金项目(22B520039)。
关键词
对抗性分类博弈
代价敏感学习
朴素贝叶斯
预期效用
整数线性规划
入侵检测
假阳性/假阴性
adversarial classification game
cost sensitive learning
Naive Bayes
expected utility
integer linear programming
intrusion detection
false positive/false negative
