Differentially Private Dimensionality Reduction via Dual-Generator Framework

In the era of big data, the conflict between data privacy protection and sharing utilization has become increasingly prominent. This paper proposes a dimensionality reduction differential privacy method based on a dual-generator framework. By pretraining a generator on public datasets to obtain low-dimensional latent representations, and then training a private generator with differential privacy mechanisms in the low-dimensional space, the method effectively mitigates noise accumulation and mode collapse in high-dimensional scenarios. A Wasserstein GAN with gradient penalty is employed to optimize gradient clipping, enhancing the quality of generated data. Experiments on the MNIST dataset demonstrate that, compared to traditional DP-GAN, the proposed method achieves significant improvements in metrics such as Fréchet Inception Distance (FID) and classification accuracy, striking a better balance between privacy protection and data utility.