Federated learning faces critical challenges in maintaining model performance under non-IID data distributions and mitigating security risks such as privacy leakage from gradient attacks or malicious clients in heterogeneous environments. To address these issues, this paper proposes PFLUG, a personalized federated learning scheme that integrates dynamic user grouping and group signatures. For data heterogeneity, PFLUG designs a dynamic client grouping mechanism based on Jensen-Shannon divergence and feature similarity, enabling hierarchical clustering that adaptively captures evolving data distributions. A dual-path model decoupling architecture separates global feature transformation layers from personalized embedding layers, harmonizing generalization and specificity through hierarchical aggregation and adaptive weight fusion. To ensure privacy, PFLUG employs an enhanced group signature protocol based on the Boneh-Boyen-Shacham scheme, which supports verifiable anonymity, traceable accountability, and efficient authentication without exposing client identities, thereby minimizing exposure risks during frequent model interactions. While ensuring security guarantees, extensive experiments on MNIST, FMNIST, and CIFAR-10 datasets demonstrate PFLUG's superiority over state-of-the-art methods, showing that the model accuracies are 1% ~ 10% higher under severe heterogeneity.