Analysis of Intelligent Privacy Protection Mechanisms for Smart Healthcare in the Context of the Internet of Things

Abstract

In Internet of Things-based smart healthcare, while medical data sharing and model training are driving industry development, sensitive patient information is easily stolen and misused, and traditional data processing models struggle to balance security and usability. This research aims to construct an intelligent privacy protection mechanism that balances privacy protection, model accuracy, and training efficiency. A two-layer framework of "foundational reinforcement and collaborative optimization" is constructed. First, a differentiable Siamese network architecture is proposed, combined with data distillation technology. Through data distillation and reinforcement, a solid foundation for privacy protection and handling data heterogeneity is established. Building on this foundation, to further improve both privacy and model performance, a personalized federated learning method based on similarity clustering is proposed. Through client-side clustering and collaborative mechanisms, the generalization ability and overall efficiency of the model are significantly enhanced on a robust underlying foundation. Experimental results show that on the ChestX-ray14 and Medical MNIST medical imaging datasets, the federated learning method combining a differentiable Siamese network architecture with data distillation technology achieves an accuracy of 88.4% (after 60 rounds of training), with a privacy leakage probability of 14.5% and an anomaly data tolerance rate of 85.7% in the comprehensive IoT medical scenario. The personalized federated learning method based on similarity clustering has a communication overhead of 115.4 MB in 25client scenarios and an accuracy of 87% in non-IID scenarios. This research demonstrates that the designed federated learning method can effectively overcome traditional technical bottlenecks and provide a feasible path for secure data collaboration in smart healthcare.