A Transformer Approach to Log-Based Anomaly Detection

This study introduces a transformer-based anomaly detection model designed to flexibly analyze log data using semantic, sequential, and temporal features. Unlike previous methods that rely on fixed log sequence lengths, this approach adapts to varying input sizes while revealing that event occurrence plays the most critical role in detecting anomalies. Experiments across multiple datasets show stable and competitive performance, highlighting both the strengths and limitations of current benchmarks and the need for more diverse anomaly datasets.