Dynamic Decision Boundary Calibration via Nelder-Mead Optimization in a Parallel CNN-BiLSTM Spatial-Temporal Hybrid Architecture for Robust DDoS Mitigation

Abstract

Modern cloud infrastructures and multi-tenant perimeter networks face severe operational risks from high-velocity volumetric anomalies and asymmetric traffic skews. Traditional signature-based systems and static deep learning models frequently suffer from line-rate processing bottlenecks, alert fatigue, and poor adaptation to zero-day attack variants. To address these perimeter protection challenges, this study presents an edge-deployable, spatial-temporal defense architecture integrating a parallel Conv1D–BiLSTM feature extraction core with dynamic Nelder-Mead decision boundary calibration. The system processes streaming flow telemetry by concurrently mapping localized packet-level signatures and tracking long-term chronological sequence dynamics, bypassing the temporal context degradation typical of standard sequential models. Evaluated across distinct enterprise and cloud environments—the UNSW-NB15 benchmark dataset (39 features) and the high-density BCCC-cPacket-Cloud-DDoS-2024 profile (50 features)—the system achieves global classification accuracies of 95.00% ( ) and 97.00% ( ), respectively. Parametric Student's paired -tests, Wilcoxon signed-rank evaluations, and post-hoc Tukey HSD analyses ( ) confirm that these detection gains stem from architectural integration rather than stochastic initialization. With a memory footprint of 2.45 MB and an average inference latency of 30.77 ms per 10,000-packet batch, the proposed framework provides a practical, high-throughput solution for real-time cyber telemetry isolation in modern cloud networks.