不均衡数据集下的入侵检测

在直推式支持向量机(transductive support vector machine, TSVM)中,迭代过程中样本标注错误会导致错误传递,影响下一次迭代中样本标注准确度,使得错误不断地被积累,造成最终分类超平面的偏移。在不均衡数据集下,传统支持向量机(support vector machine, SVM)对样本分类的错误率较高,导致TSVM在每次迭代中标注样本准确度不高。针对此,本文提出一种不均衡数据集下的直推式学习算法,该算法依据各类支持向量的密度分布关系动态计算各类的惩罚因子,提高每次迭代中样本标注的准确度,算法在继承渐进赋值和动态调整规则的基础上,减少分类超平面的偏移。最后,在KDD CUP99数据集上的仿真实验结果表明该算法能够提高TSVM在不均衡数据下的分类性能,降低误警率和漏报率。

Intrusion detection on imbalanced dataset

In transductive support vector machine, sample labeling error will result in error propagation in the iterative process. It affects the accuracy of sample labeling in the next iteration and makes mistakes constantly being accumulated. Eventually leading to classification hyperplane offset. Under imbalanced dataset, there is higher classification error rate of traditional SVM that causes the labeling error rate in each iterative for TSVM. Therefore, the algorithm of TSVM for imbalanced dataset is proposed in this paper. We dynamic calculates the penalty factor of every class according to the relationship of sample density of every class to improve the accuracy of labeling sample in each iterative. The algorithm inherits its rules of progressive labeling and dynamic adjusting, and reduces the offset of the classification hyperplane. Finally, experiment results with KDD CUP99 dataset show the algorithm can improve the classification performance at imbalanced dataset, especially for the minority class samples.