Abstract: To prevent attackers from restoring the training dataset and to protect the input image during the application process of deep learning image classification model, a differential privacy classification algorithm based on LRP (Layer-wise Relevance Propagation) is proposed in the paper. The relevance between image features is firstly quantified according to LRP in the proposed algorithm, then adaptive noise is added to the loss function based on feature relevance and Adam mechanism is used for model optimization. Finally, a differential private transform layer is constructed to perturb the input image and privacy budget is assigned according to feature relevance during model application stage. Experimental results show that the proposed algorithm achieves high classification accuracy in condition of privacy preservation.