Scholarlink

<p>The aim of log anomaly detection is to accurately identify anomalies in system logs, with the objective of ensuring system reliability and stability, thereby mitigating avoidable losses. While anomaly detection schemes in this field have achieved some success, prior research typically relies solely on specific anomaly data for model training. This practice often fails to encompass all potential anomaly types, thereby limiting the generalizability of models in practical applications. Furthermore, the high dimensionality and dynamic nature of log data often pose challenges for traditional anomaly detection methods in effectively addressing novel or unknown anomaly patterns. Consequently, this paper introduces LogAnomEX, a novel unsupervised log anomaly detection model. LogAnomEX integrates a difficulty prediction module and a gated linear neural network, built upon the Electra model, to enhance its capacity in identifying unknown anomalies. The model achieves this by learning from normal log data and iteratively generating pseudo-anomalies resembling genuine anomalous logs. We evaluate LogAnomEX’s performance on the BGL, HDFS, and Thunderbird datasets, validating its effectiveness and superiority through comprehensive experimentation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LogAnomEX: An Unsupervised Log Anomaly Detection Method Based on Electra-DP and Gated Bilinear Neural Networks

  • 期刊论文

  • 发表日期：

    2025年2月20日

查看全文
  • Keyuan Qiu,
  • Yingjie Zhang,
  • Yiqiang Feng,
  • Feng Chen

摘要

The aim of log anomaly detection is to accurately identify anomalies in system logs, with the objective of ensuring system reliability and stability, thereby mitigating avoidable losses. While anomaly detection schemes in this field have achieved some success, prior research typically relies solely on specific anomaly data for model training. This practice often fails to encompass all potential anomaly types, thereby limiting the generalizability of models in practical applications. Furthermore, the high dimensionality and dynamic nature of log data often pose challenges for traditional anomaly detection methods in effectively addressing novel or unknown anomaly patterns. Consequently, this paper introduces LogAnomEX, a novel unsupervised log anomaly detection model. LogAnomEX integrates a difficulty prediction module and a gated linear neural network, built upon the Electra model, to enhance its capacity in identifying unknown anomalies. The model achieves this by learning from normal log data and iteratively generating pseudo-anomalies resembling genuine anomalous logs. We evaluate LogAnomEX’s performance on the BGL, HDFS, and Thunderbird datasets, validating its effectiveness and superiority through comprehensive experimentation.