A Switch-Based Framework for Efficient Anomaly Detection in Noisy Log Sequences

Syslog data generated by software systems constitutes a critical source for system anomaly detection. However, existing anomaly detection approaches are highly susceptible to log parsing errors. In addition, sequence models based on a single architecture exhibit limited capacity to capture complex temporal dependencies in log sequences, resulting in suboptimal detection accuracy. In this work, ParseRobustLog is proposed as an efficient log sequence anomaly detection framework based on a Switch network. It incorporates a contrastive learning strategy to align encoded representations with target log template features, thereby mitigating the adverse impact of parsing errors. Furthermore, the Switch network dynamically integrates multiple sequence models and adaptively selects the most suitable model for each log sequence, significantly enhancing detection performance, particularly on datasets with sparse log events. Experiments on multiple widely used datasets demonstrate that ParseRobustLog consistently outperforms existing state-of-the-art anomaly detection methods.