GNN-driven dynamic access control for context-embedded neighborhood fusion

Addressing the limitations of dynamic access control in cloud computing and large-scale distributed systems with respect to conceptual drift, extreme data imbalance, and reuse of established rules, we propose a hierarchical dynamic access control approach. First, we design a log-driven ABAC rule extraction method, which builds a user–operation–resource graph, applies attention-based embedding learning to detect anomalies, and converts them into explicit rules through prefix grouping and pruning. Second, we propose Hybrid-P, an online–offline fusion method that combines offline probability and online neighbor context to make real-time decisions without retraining. Third, we propose a context-aware multi-layer decision mechanism that adds a three-level threshold to the usual binary scheme, thereby balancing availability and safety. Experimental results on BGL, HDFS, Hadoop, Spirit, and Thunderbird show an average 21% gain in ROC-AUC over each data set’s best classical baseline, PR-AUC improves by more than 3 times in extreme imbalance scenarios, per-request latency less than 50 ms, and the weighted structural complexity of the generated rules remains within 23.56-103.41. Consequently, our method ensures maintainable, adaptive, and real-time access control.