In the field of artificial intelligence, high-quality data annotation is the key to improving model performance. However, issues such as insufficient local data annotation, data privacy and security and high costs of manual annotation severely restrict the training and optimization of models. In this study, we aim to address the problem of insufficient local data annotation and improved accuracy of local models in real-world smart healthcare.
Our model supports knowledge transfer across isolated datasets and allows for incremental updates from multiple teacher models. By aggregating the knowledge of multiple teacher models through a voting mechanism, it enhances the robustness and scalability of the model. This paper introduces a random noise perturbation mechanism to effectively balance the relationship between privacy protection and data utility. This ensures that attackers cannot extract the transferred knowledge by attacking the student model or the student-side server, thereby protecting the privacy of the original data and model parameters of the teacher models.
The experimental results indicate that the proposed method outperforms traditional approaches in effectiveness. Multi-party semi-supervised knowledge transfer (MSSKT) not only innovatively combines privacy-preserving techniques with knowledge transfer in a multi-party context but also demonstrates extensive applicability in practical applications, especially in fields with high privacy requirements such as healthcare.
The novelty of this paper lies in exploring how a multi-party collaborative knowledge transfer model can be utilized to achieve privacy preservation. This approach has been relatively unexplored in depth in the field of privacy-enhancing technologies and is therefore a unique contribution to the literature. The paper not only presents the theoretical framework of the model but also demonstrates its effectiveness in practice through empirical research or case studies.
