Machine learning has been widely applied in various fields, including image and speech recognition, natural language processing, and data synthesis. Generative Adversarial Networks are a type of deep learning model designed to generate data that closely resembles real-world data. GANs are widely used in areas like image synthesis, video game development, healthcare, and data augmentation due to their ability to create realistic and varied data. However, GANs also raise privacy concerns, as they may inadvertently expose sensitive information from their training data, sparking discussions around ethical use and privacy-preserving methods in AI. To deal membership inference attacks in conditional GANs, which generate both synthetic data and corresponding labels, we introduce three new models that achieve acceptable accuracy while preserving privacy. These models employ an adaptive perturbation and clipping technique, further improved through the use of normalization and dropout. We employ the Renyi differential privacy accountant to monitor privacy loss and create synthetic samples along with its matching labels. Our evaluations demonstrate that these models can produce highly realistic and accurate synthetic data, even under strict privacy protection. On average, attack accuracy is reduced by 0.01 compared to similar methods.
Ekramifard et al. (Wed,) studied this question.