Balanced and Auxiliary-Enhanced Independent Encoding Framework for Multimodal Fake News Detection

Balanced and Auxiliary-Enhanced Independent Encoding Framework for Multimodal Fake News Detection

연구 분야: Safety

학회: International Conference on Neural Information Processing

The increasing prevalence of multimodal content on social media platforms presents a complex challenge to the detection of fake news. Most existing methods utilize complex cross-modal fusion frameworks to extract multidimensional information from news content. However, these methods often rely excessively on increasing model capacity to fit the data, neglecting the unimodal distribution anomalies in the news. We propose the Balanced and Auxiliary-Enhanced Independent Encoding Framework (BAE-IE), a multi-encoder detached architecture. This framework introduces sub-supervision tasks with a restricted sphere of influence and a cross-modal consistency task to enhance the independent feature learning capabilities of the encoders. Moreover, a multi-encoder gradient balancing mechanism is proposed to guide the encoders to reach their optimal state at similar training phases. The encoder features are then attentively weighted with different insights to obtain the final judgment of a particular news sample. Extensive experiments on two widely used social media datasets for fake news detection demonstrate that BAE-IE outperforms existing state-of-the-art methods.