ATHLETE HEALTH DATA MANAGEMENT FROM THE PERSPECTIVE OF PRIVACY PROTECTION
Abstract
In response to the problems of insufficient privacy protection and model performance in traditional athlete health data analysis models, this paper explored distributed model training methods based on the framework of federated learning. The paper first divided athlete data into time segmentation, body index segmentation, sports item segmentation, and environmental condition segmentation, and used transport layer security protocols and homomorphic encryption to protect data computation. When training the local model, a lightweight decision tree was selected for training; the dynamic weighted learning was used to aggregate the model; finally, differential privacy technology was applied to protect data privacy by adding Gaussian noise, and some optimization methods were used to improve model performance. In the third experiment of model performance, the precision of the model in this paper reached 98.45%. This indicated that the model had extremely high accuracy and reliability in classification and regression tasks. In the speedup ratio experiment, when the synchronization interval was 50 and the number of clients was 200, the speedup ratio of the model in this paper was 4.01, reflecting that the model can effectively improve training efficiency with the participation of multiple clients. In the privacy leakage risk test, the success rate of the model in this paper was the lowest under attack, at 1.5%, 2.3%, and 1.4%, respectively. Finally, in the model loss test, the model in this paper experienced the fastest decline in the initial stage, with a final convergence value of 0.3, which was the smallest among the tested models. The data showed that the model studied in this paper had good performance and privacy protection ability.