AI-Based Cyber Threat Detection: A Synthesis of Deep Learning Approaches and a Layered Detection Framework

Cyber threat detection must address high-volume telemetry, evolving attack behaviour, and the limitations of systems that depend exclusively on previously observed signatures. This paper presents a qualitative synthesis of three survey studies on deep learning-based cybersecurity, covering neural network architectures, the end-to-end intrusion detection pipeline, and broader cybersecurity applications. The comparative analysis identifies key challenges, including dataset representativeness, generalization to unseen attacks, false alarms, explainability, adversarial robustness, and deployment costs. Based on these findings, a five-layer conceptual framework is proposed: data collection and storage, preprocessing and representation, detection engine, alerting and investigation, and feedback with controlled model maintenance. Privacy, explainability, and resource efficiency are considered cross-cutting requirements. The framework is a conceptual design proposal and has not been implemented or experimentally validated. Future work should focus on reproducible implementation, leakage-resistant evaluation, baseline comparisons, and operational performance measurement.