Порівняльний аналіз методів виявлення аномалій у логах програмного забезпечення

Abstract

This study evaluates classical (PCA, Isolation Forest) and deep learning (DeepLog, LogAnomaly, LogBERT) models for log anomaly detection using HDFS and REST API datasets. Results show that model effectiveness strictly depends on the anomaly type. Classical models efficiently identify volumetric deviations, such as frequency and latency, via statistical outliers. Conversely, deep learning architectures, notably LogBERT, excel at detecting complex semantic and sequential logic violations by capturing system context. However, high computational costs limit neural networks in real-time monitoring, advocating for a hybrid approach.