Design of Federated Learning and Data Sharing System for Environmental Information System in Multi-Campus Environment

In multi-campus environmental information systems, data silos and privacy risks caused by raw-data sharing hinder collaboration. Similar challenges appear in distributed engineering monitoring, including electromagnetic environmental sensing, where multiple sites must share knowledge without exposing local data or infrastructure details. This paper designs a cross-campus collaborative modeling system using federated learning and differential privacy. The system adopts a hierarchical federated architecture: campuses train models locally and upload only model parameters with calibrated noise to a central server. Dynamic weighting and contribution evaluation mechanisms are integrated to improve generalization under non-IID data distribution while preserving efficient, privacy-protected data collaboration. A hybrid neural network combining one-dimensional convolutional neural networks, gated recurrent units, and an attention mechanism is constructed for environmental time-series characteristics, and gradient-level differential privacy is applied. Experiments using air-quality data from four campuses show that the federated model achieves an 18.1% lower mean absolute error for PM2.5 prediction than independent models. Communication overhead is reduced by 99.1%, and in extreme non-IID scenarios, the error is further reduced by 9.9% compared with standard FedAvg. The results demonstrate improved collaborative modeling for distributed environmental data and provide a transferable architecture for privacy-preserving sensor networks, including electromagnetic monitoring and wireless infrastructure assessment.