Federated Learning (FL), as a distributed machine learning paradigm, has been predominantly employed in recent years for training global models. Stochastic Gradient Descent (SGD) is widely adopted in FL as a result of its strong performance. However, FL faces inherent privacy leakage risks. Differential Privacy (DP), a rigorous mathematical framework designed to safeguard data security by injecting noise into gradients, has been extensively incorporated into FL systems. Nevertheless, the introduction of DP often leads to degraded convergence in FL models. Furthermore, most existing DP-based FL studies inadequately address data heterogeneity or operate under the assumption of a fully trustworthy server while overlooking significant disparities in client data batches and scales. Given that Fisher Information can precisely quantify parameter sensitivity, thus reflecting their relative importance, this paper proposes F-LDP, a Fisher-enhanced Personalized Heterogeneous Federated Learning algorithm with differential privacy, to address these challenges. Our algorithm accurately estimates the true noise level in the model, significantly improving FL convergence while ensuring robust privacy guarantees. By leveraging hierarchical Fisher Information to evaluate the informational value of parameters, selectively preserving high-value local parameters during training. These parameters, deemed critical, are protected from excessive noise interference. To validate the efficacy and robustness of F-LDP, we performed extensive experiments in multiple data sets and compared the results with baseline methods. Empirical results demonstrate that F-LDP superior performance compared to existing DPFL methods.