Protecting children on the internet using deep generative adversarial networks

In this paper, to control children's access to malicious information on the internet, a data sanitisation method based on deep generative adversarial networks is proposed. According to the proposed approach, an autoencoder inside the generator block by adding some noise implements the transformation of sensitive attributes considered dangerous for children, and the logistic regression inside the discriminator block performs the classification of the transformed data. To maintain the usefulness of the information during data transformation, the privacy and utility rates of the sanitised data are measured in terms of expected risk, and the optimal consensus between these two parameters is achieved by applying the minimax algorithm. In the experiments, the classification algorithm has recognised the class of sensitive data with low accuracy, and the class of non-sensitive data with high accuracy. © 2020 Inderscience Enterprises Ltd.