Differential Privacy Generative Adversarial Networks Based on Dynamic Learning Rate Constrained Adaptive Gradient Algorithm

The Generative Adversarial Network (GAN) based on differential privacy(DP) aims to protect data privacy through generative models. However, it faces challenges in balancing the trade-off between generation quality and privacy protection, as well as ensuring stable training. This paper proposes a novel algorithm called Adaptive Bound Differential Privacy GAN (ABDP-GAN), which applies the AdaBound optimizer to the training of differentially private GANs. By dynamically adjusting the learning rate, the algorithm accelerates model convergence, smooths out the impact of differential privacy noise, and ensures stable training. Experimental comparisons with other algorithms show that ABDP-GAN not only maintains the utility and privacy of image data but also achieves more stable training and higher convergence efficiency.