Integrating sophisticated, high-dimensional deep learning architectures into modern multi-cloud enterprise frameworks has vastly improved the precision of automated threat identification. Nonetheless, these convoluted, multi-layered analytical structures naturally operate as opaque mechanisms, creating significant validation and trust challenges for network security operations. This paper presents a systematic literature survey exploring the paradigm transition from legacy, shallow machine learning models to deep temporal configurations evaluated against standard benchmark repositories (NSL-KDD, CICIDS, and UNSW-NB15). Furthermore, it provides an in-depth analysis of contemporary, post-hoc Explainable Artificial Intelligence (XAI) integration paradigms—specifically emphasizing SHAP and LIME frameworks—engineered to balance the tension between predictive optimization and interpretability within production infrastructures.