Temporal Logical Attention Network for Log-Based Anomaly Detection in Distributed Systems.

Area of Science:

• Computer Science
• Artificial Intelligence
• Systems Engineering

Background:

• Anomaly detection in distributed systems is complex due to temporal dependencies, diverse states, and causal relationships.
• Log analysis is crucial but challenging for identifying subtle anomalies across components.

Purpose of the Study:

• To introduce a novel deep learning framework, the Temporal Logical Attention Network (TLAN), for robust anomaly detection in distributed system logs.
• To address the limitations of existing methods in capturing complex temporal and logical patterns within log data.

Main Methods:

• Developed a Temporal Logical Attention mechanism to model time-series patterns and logical dependencies across distributed components.
• Implemented a multi-scale feature extraction module to capture system behaviors at various temporal granularities.
• Introduced an adaptive threshold strategy for dynamic detection sensitivity adjustment based on system load and interactions.

Main Results:

• TLAN achieved a 9.4% improvement in F1-score and a 15.3% reduction in false alarms compared to existing methods.
• Demonstrated low latency for real-time anomaly detection.
• Showcased effectiveness in identifying complex anomalies involving multiple components and cascading failures.

Conclusions:

• TLAN effectively captures temporal patterns and logical correlations in log sequences, making it suitable for modern distributed architectures.
• The framework exhibits strong generalization capabilities across different system scales and deployment scenarios.
• TLAN offers a significant advancement in distributed system log anomaly detection.