In real-world signature verification scenarios, signers (users) are typically distributed across multiple clients. To ensure user privacy while improving training efficiency and model performance, clients commonly adopt distributed federated training to optimize local model functions. However, in the distributed federated learning environment for handwritten signature verification models, clients often encounter premature overfitting issues caused by insufficient local training data, which adversely affects the generalization ability and learning efficiency of feature extractors. To address this problem, this paper proposes a novel personalized federated learning algorithm named Multi-Head Frequent Cross-Client Collaborative Parameter Iteration (MH-FCCPI). The proposed method integrates three key components: a high-frequency communication strategy, a multi-head training mechanism, and a parameter aggregation method specifically designed for the feature extraction layer. During the training phase, the local model employs both a user classification head and a genuine/forgery signature classification head. Moreover, iterative collaboration is achieved through the aggregation of personalized feature extractor parameters, thereby enhancing the efficiency of model parameter optimization. Extensive experiments conducted on three benchmark datasets (GPDS-10000, BHSig260-Bengali, and BHSig260-Hindi) demonstrate that the proposed framework significantly mitigates overfitting issues, effectively addresses data heterogeneity in real-world scenarios, and outperforms state-of-the-art federated learning methods in terms of both computational efficiency and generalization capability.