The increasing adoption of IoT devices has led to the generation of vast amounts of distributed and heterogeneous data.Such data are increasingly complex, encompassing multiple sub-concepts and categories. However, existing methods struggle to handle such data while safeguarding client-sensitive information.To this end, this paper proposes a novel framework, FMIML, for federated multi-instance multi-label learning, which combines the strengths of federated learning (FL) and multi-instance multi-label learning (MIML).In this framework, we propose two adaptive aggregation methods—label richness and label balance—to better aggregate client models.These methods dynamically adjust aggregation weights based on the label distributions of client data.Experiments on four benchmark datasets demonstrate that the proposed methods consistently outperform traditional FL methods.