From Task-Specific Learning to Network-Native Intelligence: A Comprehensive Review of Machine Learning and Artificial Intelligence in Modern Networks

Machine learning (ML) and artificial intelligence (AI) are no longer peripheral optimization tools for networking; they are becoming integral to how modern networks are measured, controlled, secured, and evolved. Yet the literature remains fragmented. Existing surveys usually focus on one sub-domain at a time—for example encrypted traffic analysis, data-center networking, routing, edge intelligence, or 6G—and therefore under-emphasize the deeper shift from task-specific models to network-native intelligence. This review synthesizes recent literature from roughly 2020 to early 2026, with emphasis on the 2021–2025 wave, and organizes the field through four coupled axes: network lifecycle, deployment scope, learning paradigm, and operational constraints. We examine how supervised, self-supervised, graph-based, reinforcement, federated, generative, and foundation-model approaches have been used for traffic analysis, anomaly and intrusion detection, routing and congestion control, resource orchestration , edge/cloud/data-center optimization, and AI-native mobile/6G systems. We then compare representative studies along data assumptions, generalization behavior, online adaptability, interpretability, systems cost, and reproducibility. Our central argument is that the next phase of AI for networking is not simply “more powerful models” but closed-loop, network-native intelligence: systems that unify perception, reasoning, decision, verification, and actuation under realistic constraints such as privacy, energy, latency, safety, and cross-domain inter-operability. Based on this synthesis, we identify the main review gap in current literature: the lack of a unified, deployment-aware, lifecycle-centric perspective that spans from packet/flow analytics to autonomous network operation and emerging foundation models. We conclude with a concrete research agenda covering trustworthy online learning, digital twins, synthetic data, domain-adapted 1 foundation models, multi-agent control, and sustainable AI for communication networks.