A New Method of the Convolutional Neural Network Structure Improvement to Protect Data Based on the Obfuscation

This paper proposes a new method to improve the structure of convolutional neural networks to preserve training data based on obfuscation technique. The purpose of this approach is to protect data in a distributed environment when parties want to share datasets for training while ensuring data privacy. To solve this problem, we use a technique that obfuscates the input feature matrix in each region with the size of a pooling window randomly. Based on the properties of the max pooling function and the permutation of positions in each filter window, the resulting matrix has a constant result. The proposed method could be applied to all convolutional neural network models by editing the first convolutional layer. The input matrix shuffling is done randomly so it completely protects your privacy without changing the accuracy. The proposed method has been tested according to K-fold cross validation, achieving an average accuracy of 99.11% with an average error of deletion of 0.0882%.