The deployment of high-throughput deep neural networks within modern enterprise multi-cloud backbones has significantly advanced the accuracy of automated anomaly tracking. However, their highly complex, multi-layered topologies operate as opaque black boxes, creating substantial validation and trust barriers for security operations teams. This paper provides a comprehensive literature survey analyzing the structural shift from traditional shallow machine learning classifiers to deep temporal topologies using benchmark corpuses (NSL-KDD, CICIDS, and UNSW-NB15). Furthermore, it reviews contemporary post-hoc Explainable AI (XAI) integration paradigms, focusing on SHAP and LIME architectures designed to manage the performance-trust trade-off across production boundaries. We provide a rigorous analysis of classification metrics, mathematical foundations of feature attribution, and practical implications for next-generation security operations centers.