In response to the data transmission security risks faced by enterprise-level financial and budgeting systems in open and interconnected environments, this paper proposes an anomaly detection framework, ZT-GSAD, which integrates zero-trust architecture with graph neural networks. At the system level, the framework adheres to the zero-trust principles of “never trust, always verify, and least privilege,” introducing continuous identity authentication and lightweight hash auditing to construct a traceable and verifiable data processing chain. At the methodological level, GraphSAGE is employed to perform structural modeling of multi-entity relationships among accounts, projects, and suppliers. Through the incorporation of residual connections and DropEdge mechanisms, the framework enhances robustness and amplifies “neighborhood inconsistency” signals to effectively characterize potential anomalies. Experimental results on the Credit Card Fraud Detection Dataset 2023 demonstrate that ZT-GSAD achieves superior performance—accuracy (99.78%), precision (99.58%), recall (99.99%), and F1-score (99.79%)—compared with multiple baseline models, verifying its strong discriminative capability and robustness in high-dimensional, nonlinear financial scenarios. These findings indicate that the proposed framework achieves a unified representation of “identity trust” and “behavioral trust” under the zero-trust paradigm, providing an interpretable, verifiable, and practically feasible solution for constructing intelligent financial risk control systems.
Chen et al. (Thu,) studied this question.