LogCDM: Interpretable Log Anomaly Detection via Behavioral Graph Drift and Adaptive Signal Fusion
arXiv cs.AI · 10 Nov 2026, 03:48 AEST
LogCDM is a compact log anomaly detection model that operates without the need for GPUs at inference time. It achieves high F1 scores while utilizing three complementary signals for scoring log sessions.
Key facts
- Log-based anomaly detection is increasingly dominated by large pretrained models, from BERT-scale encoders to 8B-parameter language models, which require GPUs at inference time and are difficult to deploy on the monitoring nodes where logs are produced.
- We present LogCDM, a compact detector that scores each log session using three complementary signals: the next-event surprise of a small causal Transformer, the distance between a graph-attention embedding of the session's own transition graph and that of a normal reference graph, and a transition-drift statistic that we show is exactly a frequency-weighted chi-squared divergence between transition kernels.
- With 8.1-8.8M parameters and no GPU required at inference, LogCDM reaches F1 scores of 0.984 on HDFS, 0.870 on chronologically split BGL, and 0.850 on Thunderbird, within 0.013-0.12 of an 8.1B-parameter LLM detector evaluated under the same protocol and above all compact baselines on BGL.
- It scores a session in 5.5-10.5 ms on a CPU.
- Ablations show that the structural signals add 0.025-0.030 F1 on BGL and Thunderbird but almost nothing on HDFS, consistent with the learned fusion weights.
Read aloud
LogCDM is a compact log anomaly detection model that operates without the need for G P U s at inference time. It achieves high F1 scores while utilizing three complementary signals for scoring log sessions. Log-based anomaly detection is increasingly dominated by large pretrained models, from BERT-scale encoders to 8B-parameter language models, which require G P U s at inference time and are difficult to deploy on the monitoring nodes where logs are produced. We present LogCDM, a compact detector that scores each log session using three complementary signals: the next-event surprise of a small causal Transformer, the distance between a graph-attention embedding of the session's own transition graph and that of a normal reference graph, and a transition-drift statistic that we show is exactly a frequency-weighted chi-squared divergence between transition kernels. With eight point one-eight.8M parameters and no G P U required at inference, LogCDM reaches F1 scores of zero point nine eight four on HDFS, zero point eight seven zero on chronologically split BGL, and zero point eight five zero on Thunderbird, within zero point zero one three-zero point one two of an eight.1B-parameter L L M detector evaluated under the same protocol and above all compact baselines on BGL. It scores a session in five point five-ten point five ms on a CPU. Ablations show that the structural signals add zero point zero two five-zero point zero three zero F1 on BGL and Thunderbird but almost nothing on HDFS, consistent with the learned fusion weights.
Sources in this cluster
- LogCDM: Interpretable Log Anomaly Detection via Behavioral Graph Drift and Adaptive Signal Fusion
arXiv cs.AI · 10 Nov 2026, 03:48 AEST