Range aggregation serves as a fundamental operator in spatial analytics, yet faces critical challenges in federated ecosystems where autonomous data silos restrict raw data sharing and exhibit heterogeneous distribution characteristics. To address these limitations, we propose ASFRA-DP, a deep learning-driven acceleration engine tailored for federated spatial data. By integrating probabilistic graphical models with Monte Carlo approximation, ASFRA-DP achieves efficient processing of circular range queries while preserving data locality. Moreover, we design Density-Aware DP, a privacy-utility optimization mechanism that dynamically injects Laplace noise into each silo based on localized data density. This approach strictly satisfies \(\epsilon \) -differential privacy under adaptive composition guarantees, preventing cumulative leakage during repeated queries. Empirical evaluations on multi-silo benchmarks demonstrate that our framework reduces query latency by 3.8 \(\times \) compared to baseline methods, while maintaining competitive accuracy.