Residual Neural Networks and Gray Relational Analytics for Cloud-Native Security AI-Driven Multivariate Fraud Detection, Adaptive Threat Prevention, and Kubernetes Migration

Abstract

This paper introduces a hybrid cloud-native security framework that integrates Residual Neural Networks (ResNets) with Gray Relational Analysis (GRA) to enhance multivariate fraud detection and adaptive threat prevention in large-scale, containerized environments. Modern cloud infrastructures face increasing challenges from high-dimensional transactional data, rapidly evolving adversarial behaviors, and the operational complexity of Kubernetes migration. To address these issues, the proposed framework leverages GRA to quantify the relational strength between features and fraud indicators, generating interpretable feature-weight vectors that guide both shallow multivariate classifiers and deep ResNet-based architectures. These weights improve feature prioritization, reduce noise, and enhance the gradient flow during ResNet training, leading to superior precision, recall, and robustness in highly imbalanced datasets.

The system incorporates an adaptive threat-prevention module that dynamically adjusts model thresholds, cost-sensitive loss functions, and risk-level parameters based on real-time telemetry, drift detection, and behavioral indicators. To ensure scalability and operational resilience, a Kubernetes-centric migration strategy is employed, containerizing preprocessing pipelines, GRA engines, model training workflows, and inference services within zero-trust, autoscaling clusters. This approach supports secure multi-tenant isolation, CI/CD automation, and rapid rollback during model lifecycle updates. Experimental evaluations using large-scale financial transaction datasets show that the combined GRA–ResNet architecture outperforms traditional baselines by 20–35% in F1-score and reduces false positives while maintaining low-latency inference. The results demonstrate that fusing interpretable GRA-driven feature weighting with deep residual learning and Kubernetes-native deployment provides a scalable, adaptive, and production-ready solution for modern cloud security ecosystems.