FSECAD: Feature-Selected Explainable Cloud Anomaly Detection Framework

Abstract

High-dimensional telemetry data is constantly generated by modern cloud platforms, which presents serious scalability, interpretability, and real-time performance difficulties for anomaly detection. Despite the fact that ensemble-based detectors frequently attain excellent accuracy, feature redundancy, opaque decision-making, and significant computing overhead restrict their applications.
This paper introduces FSECAD (Feature-Selected Explainable Cloud Anomaly Detection), an effective and interpretable framework designed for cloud telemetry streams, to overcome these drawbacks. Compact, transparent, and production-ready anomaly detection is made possible by FSECAD’s integration of Stability-Aware Hybrid Feature Selection (SHFS) and Feature-Centric Explainable Anomaly Attribution (FCEA). By simultaneously improving relevance, redundancy, and stability across time windows, SHFS lowers the initial 41-dimensional feature space to 11 temporally stable and highly discriminative features. ration layer. In comparison to baseline approaches, experimental evaluation on typical cloud benchmarks shows a 92.8% F1-score, 67% shorter inference latency, and 73% lower memory use. All things considered, FSECAD offers a reliable and efficient solution for scalable anomaly detection in cloud settings.