HiparaLog: Improving Log-based Anomaly Detection through Parameter Feature Integration
Guangming Li, Jiaqing Mo, Gang Zhou, Cheng Li · 2024
Anomaly detection methods based on system logs are effective means of maintaining large-scale systems. In recent years, numerous excellent log anomaly detection methods have emerged. However, existing log anomaly detection methods identify anomalous logs either through the single template feature or by combining template features with the contextual information, ignoring the role of parameter value features in log messages. In this paper, we propose a novel log anomaly detection method named HiparaLog. HiparaLog converts each raw log message into a semantic vector, which incorporates both parameter features and template features. It then detects anomalies by employing a self-attention-enhanced GRU model, which has the ability to capture the global dependencies and contextual information within the log sequence. In this way, HiparaLog is able to detect abnormal logs from multiple dimensions. Our experimental results on two of the most challenging datasets demonstrate that HiparaLog achieves superior performance compared to current state-of-the-art methods.