Federated learning (FL) offers a powerful paradigm for collaborative model training while safeguarding data privacy, yet its deployment in maritime contexts, such as target detection. However, its practical application is challenged by statistical data heterogeneity, system scalability bottlenecks, and the unverifiable computational integrity of the central server. To address these challenges, we introduce PHV-FL, a personalized hierarchical verifiable federated learning scheme tailored for maritime target detection. PHV-FL integrates a hierarchical architecture to enhance scalability, and introduces the data-centric DOVWS and the model-centric MRDWS to mitigate the impact of imbalanced datasets and achieves a balance between local and global performance. To ensure trust, the framework incorporates zk-SNARK, enabling efficient, constant-time validation of the server’s aggregation. Experiments conducted on a heterogeneous maritime target detection dataset demonstrate that PHV-FL achieves a superior balance between local model performance and cross-scenario generalization while maintaining minimal verification overhead.