<ul><li>ADALog is an adaptive, unsupervised anomaly detection framework designed for analyzing heterogeneous log data in modern software systems.</li><li>The framework utilizes a transformer-based, pretrained bidirectional encoder with masked language modeling to capture syntactic and semantic patterns for accurate anomaly detection.</li><li>ADALog operates on individual unstructured logs, extracts contextual relationships, and uses adaptive thresholding on normal data to identify anomalies dynamically.</li><li>Evaluation on benchmark datasets BGL, Thunderbird, and Spirit shows ADALog's strong generalization and competitive performance compared to existing supervised and unsupervised methods.</li></ul>

ADALog: Adaptive Unsupervised Anomaly detection in Logs with Self-attention Masked Language Model

Discover more