COVID-19 Content Safety Detection Algorithm Based on Multimodal Robustness

This paper conducts a content safety analysis based on a multimodal dataset of COVID-19-related social media samples. A joint model of multimodal images and long texts is employed to perform an in-depth analysis of sample content. Relevant text, images, and other materials related to the pandemic topics on social platforms are collected to create a multimodal dataset of social media discourse on the pandemic. Traditional machine learning algorithms are used for in-depth data processing and topic analysis of the multimodal dataset, studying the topics people are concerned about, the main methods and ways of inappropriate discourse. Based on these results, discourses that lead to social insecurity are categorized into several categories. With this categorization, the dataset is labeled, and a multimodal deep learning model is designed. The model is used to conduct deep learning on the generated multimodal dataset, extracting their respective features, conducting feature fusion, and classifying the samples in the dataset. The discourses are divided into two major categories and several subcategories. Based on this, we propose the COVID-19 Content Safety Detection Algorithm (CSDMC, Content security detection algorithm of COVID-19 epidemic based on multimodal strong robustness). When compared with other algorithms in the experimental comparison, it achieves a relatively high accuracy rate. The experiment proves that this algorithm can effectively extract multimodal features and achieve a higher accuracy rate.