Development of machine learning based algorithms for anomaly detection from system logs

Today, with computer systems affecting every moment of our lives, ensuring their sustainability and security has gained great importance. To ensure this, we can benefit from the logs produced by the systems. Anomalies in the logs can be detected and necessary measures can be taken. System logs are often large in volume. Manually examining log files will take time and may delay the measures that need to be taken. There have been many studies on collecting and analyzing logs and detecting anomalies. In this study, anomaly detection using GNN was performed on HDFS and BGL datasets, which are frequently used in the literature. In addition, the performances obtained with other machine learning (ML) techniques obtained with the same datasets in the studies in the literature were compared. Results for an F-score of 0.9969 for the HDFS dataset and an F-score of 0.9999 for the BGL dataset were obtained. It has been observed that our study was more successful compared to the results of the LogGD study. GNNs are more stable in classification than other models by better capturing contextual features in system logs through nodes.