LogDTL: Network Log Template Generation with Deep Transfer Learning
LogDTL: Network Log Template Generation with Deep Transfer Learning
复制标题
DOI：
--
复制
发表时间：
2021-05
期刊：
影响因子：
--
通讯作者：
Thi-Thu-Trang Nguyen;Satoru Kobayashi;K. Fukuda
中科院分区：
文献类型：
--
作者：
Thi-Thu-Trang Nguyen;Satoru Kobayashi;K. Fukuda
分析网络日志在系统管理和维护中发挥着越来越重要的作用。因此，越来越多的新技术和新模型被提出用于日志自动分析。日志模板生成是应用这种复杂技术的关键第一步。本文提出了一种日志模板自动生成框架(LogDTL)，该框架将转移学习技术应用于深度神经网络(DTNN模型)中，以克服人工标注所生成模板的准确性和人力资源之间的权衡。我们的评估结果表明，DTNN的性能明显优于著名的监督方法(CRF)。DTNN只用一个训练样本就达到了91%的单词准确率，而CRF达到了78%的单词准确率。
Analyzing network logs is increasingly playing an essential role in system management and maintenance. As a result, more and more new techniques and models have been proposed for automatic log analysis. Log template generation is the essential first step to apply such sophisticated techniques. This article presents an automatic log template generation framework (LogDTL) in which transfer learning technique is used in the deep neural network (DTNN model) to overcome the trade-off between the accuracy of the generated template and human resources for manual labeling. Our evaluation results show that DTNN significantly outperforms a well-known supervised method (CRF). DTNN achieves 91% of word accuracy with only one training example though the CRF achieves 78% of word accuracy.