Optimization of Multi-source Government Data Fusion and Privacy Protection Mechanism Based on Federated Learning

At present, multi-source government data generally has problems such as data islands, feature heterogeneity, and privacy leakage risks in the fusion process, which seriously restricts the improvement of cross-departmental data collaboration and intelligent governance. To this end, this paper proposes a fusion and privacy protection mechanism optimization method based on horizontal federated learning and uses a differential privacy mechanism with a local adjustable privacy budget to perturb the model update parameters to enhance data privacy security. A feature alignment module driven by maximum mean difference is constructed to achieve spatial mapping and distribution unification of heterogeneous government data. Combined with communication compression and dynamic weight aggregation strategies, the stability and resource utilization of the federated training process are effectively improved. Experimental results show that under the premise of protecting privacy, the fusion accuracy of the global model reaches $\mathbf{9 2. 6 \%}$, and the MMD (Maximum Mean Discrepancy) value of the data source after feature alignment drops to $\lt 0.1$. The method in this paper shows good adaptability and optimization performance in multi-source government data fusion and privacy protection, providing a feasible path for the secure sharing of government data.