Machine-learning intrusion detectors can provide limited decision transparency in shared-server environments, where analysts must distinguish routine service activity from suspicious flows. This study proposes DIRAF (Direction-Aware Rule-Augmented Forest), a two-stage cascade comprising a conservative, interpretable rule gate and a Random Forest (RF) classifier for unresolved flows. The framework is supported by Quadruple-View Flow Engineering (QVFE), which organizes behavioral predictors into temporal, flow-morphology, port-semantics, and TCP-flag views, and by the Triangulated Attribution Protocol (TAP), which reconciles inferential, permutation-based, and SHAP feature rankings. Evaluation used the CIDDS-001 External Server Week 1 benchmark (172.838 flows); the primary binary task included 156.950 flows labeled normal or suspicious, while 15.888 unknown flows were excluded. On the held-out binary test set, DIRAF achieved 100% accuracy, precision, recall, F1-score, and AUC under the evaluated configuration. Its rule gate resolved 42,3% of test traffic, covering 61,9% of suspiciously labeled flows with 100% precision; the remaining flows were classified by the RF. The cascade reduced mean batch decision latency from 8,01 to 4,77 µs per record. These results are specific to this benchmark: direction laundering reduced Model B accuracy to 68,8%, and the direction-label association limits external validity. DIRAF therefore demonstrates an auditable cascade design on CIDDS-001 Week 1, not universal or production-level detection performance.
Copyrights © 2026