Abstract
With the rapid advancement of intelligent healthcare, the privacy protection of personal health data has become a critical issue that urgently needs to be addressed. To tackle this challenge, this work proposes a Differential Privacy-based Generative Adversarial Network for Healthcare Data (DP-GAN-HD), which integrates Generative Adversarial Networks (GANs) with differential privacy mechanisms to ensure secure data publishing. The method addresses the challenge of efficiently publishing personal health data under privacy protection. The proposed method employs a multi-generator architecture and optimizes generator parameters through gradient clipping and genetic algorithms, enhancing data privacy protection and the quality and utility of the generated data. Experimental results show that, with a privacy budget of 2.0, the accuracy of DP-GAN-HD on the Adult, Br2000, and Kaggle Cardiovascular Disease datasets reaches 0.784, 0.800, and 0.823, respectively. They all outperform other differential privacy models and are slightly lower than the real datasets, demonstrating a strong balance between privacy protection and data utility. Additionally, the model's accuracy gradually improves as the privacy budget increases. When a privacy budget is 3.0, DP-GAN-HD achieves its peak, performing nearly identically to real data. DP-GAN-HD demonstrates enhanced resistance to privacy attacks through its multi-generator framework and Gaussian noise perturbation mechanisms. These features collectively reduce privacy leakage risks while maintaining an effective balance between data utility and protection. Overall, the experimental results reveal that DP-GAN-HD excels in balancing privacy protection and data utility across different datasets, and prove its adaptability and effectiveness in intelligent healthcare applications.