Leveraging Knowledge Graphs for CheapFakes Detection: Beyond Dataset Evaluation

Abstract

The proliferation of the internet and the availability of vast amounts of information have given rise to a critical and pressing issue of fake news. Among the various forms of fake news, cheapfakes are particularly prominent in deceiving people. Existing research on cheapfakes detection has primarily focused on analyzing the context and correlation between textual and visual information, but has largely overlooked the significance of external knowledge. As a result, most previous approaches, apart from the baseline of ICME‘23 Grand Challenge on Detecting Cheapfakes, have heavily relied on evaluating the dataset itself to improve performance. However, despite achieving impressive results on public test datasets, these approaches often suffer from poor performance in real-world scenarios due to their overreliance on the given dataset. In this study, we propose a novel approach that utilizes knowledge graphs to address the issue of insufficient information from external knowledge. Unlike previous approaches, our proposal does not directly alter or participate in the public test dataset to enhance performance, which can potentially result in significant overfitting. Our proposed approach achieved an accuracy score of 83.52% on Task 1, surpassing the baseline by 1.7%, and an accuracy score of 84% on Task 2, outperforming the best result from the previous challenge by 8%.