A layer-wise Perturbation based Privacy Preserving Deep Neural Networks

Abstract

Datasets are sources of information mining where knowledge can be derived. The versatility of these dataset determines the quality of knowledge gained. However, several of these data contains personal sensitive information that can lead to infringement of privacy. Existing research tends to deliver DNN models that can preserve privacy of personal information but the accuracy of these models are rather much lower as compared to their non-privacy preserving counterparts. This is due to the degree of noise and the points where noise was added to perturb the model data. Consequently, this has led to minimal adoption of privacy preserving DNN models in the industrial world. In this paper, we present a layer-wise perturbation approach and differential privacy technique to determine points of perturbation and preserve privacy. Our approach was able to narrow down the accuracy gap between privacy-preserving and non-privacy preserving DNN model.