First Steps in Pixel Privacy: Exploring deep learning-based image enhancement against Large-scale image inference

In this paper, we present several enhancement approaches for the Pixel Privacy Task of MediaEval 2018. The goal of this task is to use image enhancement techniques to fool the state-of-the-art convolutional neural network (ConvNet) classifiers in scene classification problem, and maintain the visual appeal of images. Our proposed approaches are based on image crop, adversarial perturbations and style transfer, respectively. Firstly, we showed the potential influence of easy-to-use image processing operations, i.e., cropping (center cropping and random cropping). In perturbation-based approach, we apply a white-box technique, which makes use of the information of ConvNet classifiers. Based on the experiments, we observed some limitations of this approach, caused by, for example, image preprocessing. In addition, we demonstrated the style transfer-based approach, which was not developed for privacy protection, could be also used to reduce the effectiveness of the classifiers for Large-scale Image Inference. Specifically, we implement black-box techniques based on the Generative Adversarial Network. Experimental results showed that style transfer-based approach could address privacy protection and appeal improvement simultaneously. Copyright held by the owner/author(s).