A privacy-preserving multi-agent updating framework for self-adaptive tree model

The tree-based model is widely applied in classification and regression problems because of its interpretability. Self-adaptive forest models are proposed for adapting to dynamic environments by using active learning and online learning techniques. However, most existing self-adaptive forest models are designed under a single-agent situation. With the development of the IoT, data is distributed across multiple edge devices without geographic restrictions. A global model is trained by distributed data across multiple devices. Therefore, extending a single-agent self-adaptive forest model to a multi-agent one is useful to make the original tree-based models glow with new vitality. In a multi-agent system, the privacy-preserving problem should be addressed when sharing knowledge between agents. In this paper, we propose PMSF, a privacy-preserving multi-agent self-adaptive forest framework via federated learning. We utilize differential privacy to prevent attackers from getting the data statistics. No private data is uploaded into the server in our framework and only updated parameters are uploaded. Finally, We design local adaptation and global update procedures to ensure the ability of self-adaptation of the forest model and the ability of privacy protection in each agent, which can further improve the performance of self-adaptive forest models. To demonstrate the superiority and effectiveness of our framework, we conduct extensive experiments in an identity authentication case with two datasets.