Privacy-Preserving Remote Sensing Image Generation and Classification With Differentially Private GANs

Abstract

Generative adversarial networks (GANs) have demonstrated their remarkable capacity to learn the training data distribution and produce high-quality synthetic images, which have been widely adopted in image recognition tasks in remote sensing (RS) research communities. However, previous work has shown that using GANs does not preserve privacy, e.g., being susceptible to membership attacks, while sensitive information is vulnerable to nefarious activities. This drawback is considered severe in RS communities, in which critical researches highly value the security and privacy of the image content. Thus, to publicly share sensitive data for supporting critical researches and, in the meantime, guarantee the model accuracy trained from privacy-preserving data, this work develops GANs within the differential privacy (DP) framework and proposes an RS differentially private generative adversarial network (RS-DPGAN) for both privacy-preserving synthetic image generation and classification. Our RS-DPGAN is capable of releasing safe version of synthetic data while obtaining favorable classification results, which gives rigorous guarantees for the privacy of sensitive data and balance between the model accuracy and privacy-preserving degree. Both extensive empirical and statistical results confirm the effectiveness of our framework.