AAR-Log: A robust log anomaly detection method resisting adversarial attacks

In recent years, adversarial evasion attacks against log-based anomaly detection systems have been proven to pose severe threats. Existing detection models lack targeted defense mechanisms against such attacks, enabling malicious actors to conceal anomalous activities. This not only prevents timely detection of system failures or intrusions, but also significantly elevates security risks and potential losses. To address this challenge, this paper proposes AAR-log, a novel adversarial-resilient log anomaly detection framework. The key innovations of AAR-log include: (1) integrating various types of features including log components, log levels, and log templates to construct a more comprehensive representation of log sequences; (2) employing ensemble learning to enhance model diversity and mitigate the vulnerability of single-model approaches; and (3) incorporating adversarial training to constrain the adversarial sample space, thereby significantly improving the framework’s robustness under adversarial conditions. Extensive experiments on benchmark log datasets demonstrate that AAR-log exhibits superior adversarial robustness. Its ensemble learning mechanism and adversarial training strategy effectively enhance resistance against evasion attacks, achieving higher TPR and F1 scores compared to baseline models. Notably, AAR-log also maintains strong generalization performance in non-adversarial environment. In the HDFS dataset, AAR-log improves TPR by 2.8%–9.5% and F1-score by 4.6%–14.3% over baseline models. In the BGL dataset, it achieves even greater gains, increasing TPR by 17.1%–26.7% and F1-score by 3.9%–28.4%. These results validate the effectiveness and robustness of AAR-log in log-based anomaly detection.