Adaptive Cloud Defense Using Graph Neural Networks and LLM-Based Policy Engines
Year: 2026
Type: conference-paper
Abstract: Adaptive cloud architectures require security capabilities that are able to reason about complicated dependencies and react to changing threats. This paper introduces an adaptive cloud defense to integrate large language model-based intelligent security policy engines with the additional use of the graph neural networks. The dynamic graphs vision of system entities and interactions facilitate relational threat modeling, analysis of anomalies propagation as well as context-aware risk assessment via learned graph infusion. The policy engine uses language models to read high-level security intentions, encode regulatory and organizational requirements into implementation controls as well as to adjust enforcement plans by responding to continuous feedback loops. Bi-directional communication in between the graph inference and the policy reasoning facilitates the real time response to the changes in the topology, the mobility of the workloads, and the multi-tenant behaviors. The architecture focuses on explainability, ethical correctness, and scalability in the case of heterogeneous clouds. Simulated and benchmarking experimental trials of cloud attacks prove to be better detected, lowering response time, and uniformity in policy adherence than the static and rule-based defense. The findings show that graph intelligence can be used in conjunction with language-based policy thought to allow security.
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7203668993