论文部分内容阅读
针对传统批处理特征选择方法处理大规模骨干网数据流存在时间和空间的限制,提出基于在线特征选择(online feature selection,OFS)的网络流异常检测方法,该方法将在线思想融入线性分类模型,在特征选择过程中,首先使用在线梯度下降法更新分类器,并将其限制在L1球内,然后用截断函数控制特征选择的数量。研究结果表明,提出的方法能充分利用网络流的时序性特点,同时减少检测时间且准确率和批处理方法相近,能满足网络流异常检测的实时性要求,为网络流分类和异常检测提供一种全新的思路。
Aiming at the limitation of traditional batch feature selection method to deal with the time and space of large-scale backbone data flow, this paper proposes an online feature anomaly detection method based on online feature selection (OFS). This method incorporates online thinking into linear classification model, In the process of feature selection, the online gradient descent method is first used to update the classifier and limit it to the L1 sphere. Then the number of feature selections is controlled by a truncation function. The results show that the proposed method can make full use of the temporal characteristics of network flow, reduce the detection time and the accuracy is similar to the batch processing method, and can meet the real-time requirements of network flow anomaly detection and provide one for network flow classification and anomaly detection Kind of new ideas.