AugLog: System Log Anomaly Detection Based on Contrastive Learning and Data Augmentation

Abstract

Listen

Translate article

System logs are valuable resources for system main-tenance and troubleshooting since they record run-time status and significant events of computer systems. Detecting anomalies via system logs has been widely researched in recent years. However, current log-based anomaly detection approaches are susceptible to noise introduced by log processing and evolution. In this work, we proposed AugLog, a log-based anomaly detection method that effectively reduces the impact induced by processing and evolution noise. AugLog does not need to elaborate log parsing. After pre-processing, it leverages text data augmentation to extend the training data and simulate the unstable log data. Then it utilizes the Transformer encoders to mine the semantic information of log sequence and realizes sequence modeling with the help of contrastive loss. In this way, AugLog incorporates the log representation and anomaly detection into one step. Our evaluations on widely used log datasets indicate that AugLog achieves high performance over baseline methods. The experimental results on the unstable log dataset show that Auglog efficiently reduces the impact of noise.