Year: 2026
Type: article
Abstract: Artificial intelligence is increasingly used in cybersecurity operations for intrusion detection, anomaly discovery, malware analysis, and alert triage. However, many AI-based security systems remain weakly interpretable and disconnected from governance and risk-management practices used by security teams. This paper proposes a governance-aware explainable AI framework that transforms model outputs into analyst-readable evidence packages for threat detection and risk-based response. The framework integrates machine-learning detection, SHAP-based feature attribution, MITRE ATT&CK-oriented behavioral hypothesis mapping, contextual risk scoring, and analyst feedback. A reproducible leakage-aware validation was conducted on a public CICIDS2017-derived sample. The upstream file contained 56,661 labeled flows; 12,598 exact duplicate feature/label rows were removed before splitting, leaving 44,063 records. A group-aware held-out split produced zero exact feature-vector overlap between training and test partitions. In the binary benign-versus-attack task, sigmoid-calibrated Random Forest achieved 0.993 accuracy, 0.993 balanced accuracy, 0.995 precision, 0.990 recall, 0.993 F1-score, 0.9992 ROC-AUC, 0.9990 PR-AUC, and 0.0059 Brier score. Original-label outcomes are reported to avoid hiding rare-class behavior. The reported run used mean absolute SHAP values for global attribution and representative local SHAP examples covering true-positive, false-positive, false-negative, and true-negative cases. The risk-response layer is evaluated as an illustrative governance mechanism because the public CSV does not contain organization-specific asset criticality. Results are interpreted as leakage-aware public-sample validation, not as a full official CICIDS2017 benchmark. The framework supports transparent, reproducible, and risk-aware AI adoption in security operations centers while preserving human analyst judgment.
Source: DergiPark
Author: Erkalkan, Ercan
Institution: Marmara University
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213726371
