LogMILP: Weakly-Supervised Log Anomaly Localization via Counterfactual Perturbation
Researchers propose LogMILP, a weakly supervised framework for log anomaly detection and localization. The method uses multi-instance learning with prototype-guided structural modeling and counterfactual perturbation consistency regularization. It requires only bag-level labels to perform both bag-level detection and instance-level localization. Experiments on three public datasets show competitive performance.
Key facts
- LogMILP is a weakly supervised framework for log anomaly localization.
- It uses multi-instance learning enhanced by prototypes and perturbation.
- The method requires only bag-level labels for training.
- It performs both bag-level detection and instance-level localization.
- Counterfactual perturbation consistency regularization improves localization reliability.
- Experiments were conducted on three public datasets.
- The framework achieves competitive detection performance.
- The paper is available on arXiv with ID 2605.10988.
Entities
Institutions
- arXiv