Log Parameter Anomaly Detection System: Evaluating Accuracy with Noisy Training Data
Hironori Uchida, Keitaro Tominaga, Hideki Itai, Yujie Li, Yoshihisa Nakatoh · 2025
In system development, frequent updates often result in daily issues, increasing the demand for automated error analysis. Our research focuses on anomaly detection based on parameter combinations within text logs. Previous studies have reported high accuracy under conditions where only relevant logs were included in the training data. However, in real-world operations, it is highly likely that irrelevant noise logs will be present in the training data. Therefore, we investigated the robustness of the parameter log anomaly detection model by introducing various noise ratios. As a result, we confirmed that even with 11.1% noise log contamination in the dataset, the accuracy remained unchanged. Additionally, in comparison to previous studies, we observed improved anomaly detection accuracy for Value parameters with overlapping ranges across different combinations, achieving an F1-Score of 1.0. This improvement can be attributed to the increased variety of logs in the training data, which enhanced the performance of both the tokenizer and the AI model.