Abstract: Local differentially private federated learning (LDP-FL) provides strong privacy guarantees by adding noise at the client side, but it often struggles to balance privacy protection with model utility. To address the issues of signal-to-noise ratio (SNR) degradation caused by high-dimensional parameter perturbation and update bias in data-heterogeneous environments within LDP-FL, this paper proposes a Synergistic Differentially Private Federated Framework via Structured Filtering and Quality-Aware Aggregation (SFQA-DPFL). This framework effectively reduces parameter dimensionality by freezing general feature regions and fine-tuning restricted subspaces; simultaneously, it constructs a multidimensional reputation scheduling strategy to suppress clients that deviate from the global consensus; and it introduces a structured denoising operator to filter out high-frequency noise while preserving long-tail personalized semantic features. Extensive experiments on the Fashion-MNIST and CIFAR-10 datasets demonstrate that compared to mainstream algorithms such as PCFed-LDP, AdapLDP-FL, and FedCEO, SFQA-DPFL improves the convergence utility and accuracy of the model under privacy constraints. Furthermore, the SFQA-DPFL framework exhibits excellent performance in terms of system efficiency and communication overhead. This research provides a novel approach for optimizing aggregation mechanisms in differentially private federated learning under complex, heterogeneous environments.