Differential Privacy Image Generation with Reconstruction Loss and Noise Injection Using an Error Feedback SGD

TL;DR

This work tackles synthetic data generation under differential privacy by introducing an Error Feedback SGD (EFSGD) framework that reduces clipping bias inherent in DP-SGD. It combines reconstruction loss and targeted noise injection during generator upsampling to preserve structure and enhance diversity, all within a DP-GAN-DPAC-inspired setup that privately trains discriminators while releasing a private generator. The approach leverages Rényi DP for efficient privacy accounting and uses multiple error-tracking variables to manage the effects of gradient clipping across detector, classifier, and encoder components. Empirical results on MNIST, Fashion-MNIST, and CelebA show state-of-the-art or competitive IS, FID, and gen2real metrics under the same privacy budget, with ablations confirming the positive impact of reconstruction loss and noise injection. Overall, the method advances privacy-utility trade-offs in DP-based image synthesis for both grayscale and RGB data.

Abstract

Traditional data masking techniques such as anonymization cannot achieve the expected privacy protection while ensuring data utility for privacy-preserving machine learning. Synthetic data plays an increasingly important role as it generates a large number of training samples and prevents information leakage in real data. The existing methods suffer from the repeating trade-off processes between privacy and utility. We propose a novel framework for differential privacy generation, which employs an Error Feedback Stochastic Gradient Descent(EFSGD) method and introduces a reconstruction loss and noise injection mechanism into the training process. We generate images with higher quality and usability under the same privacy budget as the related work. Extensive experiments demonstrate the effectiveness and generalization of our proposed framework for both grayscale and RGB images. We achieve state-of-the-art results over almost all metrics on three benchmarks: MNIST, Fashion-MNIST, and CelebA.

Figures (5)

• Figure 1: Privacy-preserving data synthesis via a differential private generative model. (The left side illustrates the overall framework of our method, while the right side shows the error feedback gradient processing procedure. The algorithm updates the model parameters $x^t$ with $e^{t-1}$, injects the DP noise $w^t$, and computes the clipping error $e^t$ for the next iteration. $C_1$ and $C_2$ are the clipping thresholds for gradient clipping, respectively.)
• Figure 2: the layers of multiple noises adding in
• Figure 3: Images generated for MNIST.
• Figure 4: Images generated for Fashion-MNIST. In the penultimate column, with clear handles, the bags in our results are close to the real, and in the rest columns, the blank space at the neck of the T-shirts and the sleeves of the skirts are visible.
• Figure 5: Images generated for CelebA. There are several individuals in our results (all of these are experimental results and have not been manually selected), whereas, in the DP-GAN-DPAC, many results are even completely identical. At the same pixel level, our results excel in detail, especially in the second and third rows. Zoom in for the best view.