With the rapid development of information technology, fraudulent behaviors in multiple fields such as financial transactions, social networks, and review systems show an increasingly complex and diversified trend, which poses a serious challenge to traditional fraud detection techniques. Although current mainstream graph neural network-based methods perform well in single-agency data environments, cross-agency data sharing and collaboration are difficult due to the involvement of sensitive user information, which in turn limits the training effectiveness and generalization performance of the model. Federated learning, as an emerging privacy-preserving distributed learning paradigm, provides a feasible way for cross-agency collaborative training, but existing graph federated learning methods are mostly designed for general graph tasks, making them difficult to adapt to the class imbalance and data heterogeneity problems prevalent in fraud detection, resulting in poor performance in fraud sample identification. To address the above challenges, this study proposes a risk perception dynamic aggregation graph federated learning method (FedRPDA) for fraud detection, aiming to effectively deal with complex fraud risk event recognition across organizations. FedRPDA includes two key strategies: the typical risk dynamic aggregation strategy measures the structural risk intensity of fraudulent nodes in the client graph and combines it with a dynamic weight mapping mechanism with temporal decay characteristics to adaptively adjust the aggregation weights of clients, thus enhancing the global model’s ability to discriminate between normal samples and typical fraud samples under heterogeneous data conditions; the diversified risk average aggregation strategy integrates a variance perturbation-based feature enhancement mechanism for fraud samples with a global prototype-guided contrastive learning mechanism, which effectively improves the model’s ability to represent structurally diverse and scarce a typical fraud samples, promotes their convergence toward common anomalies in the feature space, and further enhances the model’s robustness in recognizing complex fraud risk scenarios. Experimental results on several real-world fraud detection datasets show that FedRPDA significantly outperforms existing graph federated learning baseline methods in terms of detection performance and training convergence efficiency, and demonstrates good generalization ability and practical application potential.