LogFA: Cross-system log anomaly detection based on hierarchical feature alignment and domain adversarial learning

Log anomaly detection is a pivotal technology for automated operations (AIOps) and plays a crucial role in ensuring system stability and security. System logs record the system’s operational status and other important events, and analyzing these logs helps administrators troubleshoot issues and monitor system performance. In recent years, deep learning methods, which rely on labeled samples, have been able to accurately identify anomalies in log sequences, but labeling data is costly, and anomalous samples are scarce. Furthermore, there are distribution differences in $\log$ features across different systems. To address the aforementioned issues, this paper proposes a domain-adaptive transfer learning framework for log anomaly detection, called LogFA. This framework achieves efficient cross-system knowledge transfer through a multilevel optimization strategy. Specifically, a cross-layer Maximum Mean Discrepancy (MMD) alignment method was designed to comprehensively optimize the distribution consistency of local and global features between the source and target domains, while effectively suppressing noise interference. Secondly, a subclass domain adversarial method was employed to achieve category-level fine-grained domain adaptation. Finally, the Lsoftmax loss function is introduced to optimize the classification boundary, further enhancing the robustness of the detection model. Experiments on multiple datasets, including HDFS, BGL, and Thunderbird, fully demonstrate the superior performance of $\log \mathrm{FA}$.