Exécution de notre outil de synthèse sur un nouvel article

Résumé : Logs continuously record both normal and anomalous system operational states, serving as critical data for system runtime monitoring. Log anomaly detection facilitates system debugging and root cause analysis for issues such as system failures, shutdown faults, null pointer exceptions, illegal argument exceptions, and class cast exceptions. Deep learning has been widely adopted in log anomaly detection to enhance detection accuracy. However, deep learning models require substantial labeled log data, which demands significant human effort and time investment. To address this labeling requirement, pretraining models like BERT have been introduced. However, pretraining models introduce new challenges. BERT requires finetuning of an enormous parameter set, resulting in substantial training overhead. Existing parameter-efficient strategies in log anomaly detection degrade the performance while introducing inference latency through increased model depth. Furthermore, the continuous expansion of system scale generates exponentially growing log data containing redundant log information. Current random sampling methods for training data selection fail to filter more valuable instances and exacerbate data redundancy. To address these challenges, we propose a Log Anomaly Detection based on Active Learning and Prefix (LogAL-Prefix). LogALPrefix eliminates manual prompt engineering by freezing pretraining model parameters during training and exclusively updating prefix parameters, thereby further reducing the parameter footprint of pre-training models. Additionally, LogALPrefix employs active learning strategies to select a minimal set of high-uncertainty samples to retrain the model. Experiments on the BGL dataset demonstrate that$\log A L$-Prefix can further reduce the trainable parameters compared to LogBP-LoRA.