An Improved LRP-Based Differential Privacy Preserving Deep Learning Framework
Yaling Zhang, Shibo Bai · 2021 17th International Conference on Computational Intelligence and Security (CIS) · 2021
Aiming at the privacy leakage risk of the existing differential privacy preserving deep learning framework when calculating relevance, and the problem of model training consuming too much privacy budget, this paper proposes an adaptively allocation of dynamic privacy budget based on LRP (Layer-wise Relevance Propagation) differential privacy preserving deep learning framework. In the process of the relevance calculation, noise is added to the relevance decomposition information, and the privacy budget is dynamically changed during the model training process. Theoretical analysis and experimental results show that the framework proposed in this paper can protect data privacy while maintaining a high accuracy of the model.