Deep Learning Research Based on Adaptive Differential Privacy and Gradient Security

Differential privacy theory combined with deep learning models plays a significant role in data protection. However, the rationality of privacy budget allocation directly impacts model utility. As a solution, a deep learning approach integrating adaptive differential privacy and gradient security is proposed for enhanced data protection. Initially, data features are computed using the step-by-step correlation transfer algorithm. Subsequently, Laplace noise is added based on the privacy measure of each feature. Finally, the privacy budget is allocated sensibly considering the noise protection of the average feature correlation, followed by model training with pruning operations. Evaluation in this study is based on accuracy rates and loss values, demonstrating that the performance of this algorithm matches or exceeds that of many classical algorithms across diverse datasets.