LogFree: A Low-Resource-Oriented Model for Log Anomaly Detection

As the scale of system data continues to grow exponentially, detecting anomalies from log data has become increasingly crucial. Current research in log anomaly detection relies heavily on deep learning models. However, their practical deployment is often hindered by limited computational resources. To tackle the problems of parameter inefficiency and high resource overhead in such models, this paper proposes LogFree, a lightweight model integrating word embedding, Bi-LSTM for spatiotemporal feature extraction, and multi-head attention for critical event focusing. Innovatively, layer-specific quantization reduces weight and activation precision, while neuron augmentation recovers quantization-induced accuracy loss. Evaluations on the HDFS and BGL datasets demonstrate that LogFree achieves superior overall performance compared to baseline models such as DeepLog, LogAnomaly, and LogBERT. The model significantly reduces its size, enhances training efficiency, and maintains high detection reliability even in environments with limited resources, providing a cost-effective solution for log analysis on edge devices.