The application of label differential privacy in deep learning
Chunyu Wang, Qian Yang, Zuli Wang, Zhi Sun · 2025
Differential privacy plays an important role in protecting data privacy, but it often reduces data usability and affects the classification performance of deep learning models. In this paper we propose a method based on label differential privacy for deep learning classification models. Based on the Gaussian mechanism, this paper introduces Gaussian Randomized Response Label Differential Privacy (GRR-LDP). The algorithm uses the Gaussian mechanism to generate a probability value, which is then used as a parameter for randomized response to perturb the training dataset, resulting in a perturbed set of labels for the training dataset. Experiments were conducted on the CIFAR-10 dataset and the Chest X-Ray dataset for pediatric pneumonia. Results show that under low privacy budgets, the classification accuracy on the CIFAR-10 dataset can reach 92.06%, and the ROC-AUC for classification on the Chest X-Ray dataset can reach 92.56%. GRR-LDP achieves a good balance between protecting privacy and maintaining model performance.