Privacy Preserving Method for Image Recognition based on Denoising Autoencoder

Abstract

Machine learning based on neural networks have obtained great achievements in diverse domains. Training models usually requires large, labeled datasets, which are often crowdsourced and consist of private information. There is no doubt that any private information about the datasets should not be let out by the models. To realize the goal, this article introduces a composed model based on denoising autoencoder for classification. The difference between the reconstructed images by the decoder and the actual images is then set as the optimization object while training the encoder and decoder parts, and multiple fully linked layers are added to the pretrained encoder to create a composite model for classification. The entire assembled model is then trained once more with the noised images. The constructed model produces a satisfactory result since the characteristics extracted by the encoder were what the model used to do classification. The experiment results demonstrate that the composed model can protect privacy at a low cost in model quality and accuracy compared to the baseline models which take raw images as input.