Efficient Zero-shot and Label-free Log Anomaly Detection for Resource-constrained Systems

Abstract

This paper addresses the challenge of log anomaly detection on resource-constrained edge devices. We design MaidLog, a novel framework based on large language models that achieves industry-leading zero-shot accuracy while maintaining efficiency suitable for deployment on edge devices with limited computational resources.

Key Contributions

• Designed MaidLog, an innovative LLM-based framework for zero-shot log anomaly detection
• Achieved industry-leading accuracy without requiring labeled training data
• Demonstrated superior performance across multiple real-world datasets
• Provided a practical solution for automated log analysis in edge computing scenarios

Publication Details

• Conference: 42nd IEEE International Conference on Data Engineering (ICDE 2026)
• Ranking: CCF-A, Core A*
• Year: 2026
• Status: Accepted

Authors

Zuohan Wu, Jiachuan Wang, Libin Zheng, Yongqi Zhang, Shuangyin Li, Lei Chen

Significance

This work represents a significant advancement in applying large language models to practical data engineering tasks, particularly in resource-constrained environments. The zero-shot capability makes it highly applicable to diverse systems without requiring extensive labeled data collection and training.

BibTeX

@inproceedings{wu2026maidlog,
  author       = {Zuohan Wu and
                  Jiachuan Wang and
                  Libin Zheng and
                  Yongqi Zhang and
                  Shuangyin Li and
                  Lei Chen},
  title        = {Efficient Zero-shot and Label-free Log Anomaly Detection for Resource-constrained Systems},
  booktitle    = {42nd {IEEE} International Conference on Data Engineering, {ICDE} 2026},
  year         = {2026},
  note         = {To appear}
}

More details will be added upon publication.