Automatic content moderation on social media

Abstract

Millions of users produce and consume billions of content on social media. Therefore, human-reviewed content moderation is not achievable in such volume. Automating content moderation is a scalable solution for social media platforms. In this research work, we propose an automatic content moderation pipeline based on deep neural networks. Our solution consists of two main parts: the first part classifies a given image into granular content classes; and a second part obfuscates the part of a given image that might be inappropriate for the target audience. Our proposed solution is a cost-efficient in terms of human labour and practical for deploying the real-time systems. Our classification network is trained with automatically labelled data using noise-robust techniques. Our automatic obfuscation algorithm uses the information obtained from the classification network and does not require additional annotation or supplementary training. This obfuscation algorithm presents a novel-use case of class-specific activation mappings for censoring regional explicit nudity in images. The classification network achieves a top-1 accuracy of 0.903 and a top-2 accuracy of 0.986. The obfuscation algorithm covers a minimum explicitly nude area of 0.68 on average.