Dispersed, heterogeneous entities have become the overall bottleneck for the dissemination of spatial-temporal finance data under current auditing; This issue’s core limitation is imposed by tight data privacy rules and institutional division. Traditional centralised audit models have substantial limitations in establishing continuous and immutable evidence chains for large-scale corporate consortia; thus, they often fail to ensure data sovereignty or the accuracy of anomaly detection. To overcome the shortcomings in structure mentioned above, a new decentralised analysis system based on federated learning is proposed to establish cross-institutional audit evidence chains and carry out intelligent recognitions without direct transmission of original ledgers. We develop a customised multi-party cryptographic verification (MPCV) protocol inside a decentralised deep-learning system to jointly train an anomaly-detection model across multiple nodes and keep all training data local. The Architecture introduces an adaptive gradient-aggregation mechanism that can dynamically recalibrate the institutions’ weights according to localised data density and reliability indicators; thus eliminating malicious model updates or unreliable institutions. Empirical verification of the proposed architecture uses a large-scale, proprietary data set covering cross-regional transactional Networks among prominent state-owned enterprises; specifically, it analyses the procurement and operation ledgers of municipal Metro Corporations located in the Shenzhen administrative area. The performance evaluation results show that the federated evidence-chaining model has achieved a verification accuracy of 96.7%, while also reducing the mean absolute deviation of cross-institutional abnormality detection by several magnitudes relative to single-institutional base cases. The system removes the localised forms of heuristic auditing instincts and replaces them with an absolutely valid Security Infrastructure to carry out automatic accounting overseeing and forensic analysis independently by itself.
