Uni-Detector: Towards Universal Multimodal Fake News Detection
TL;DR
The proposed Uni-Detector achieves highly competitive performance across all modality combinations on test sets, establishing the Uni-Detector as a more flexible and powerful tool for detecting fake news in real-world scenarios.
Abstract
The rapid rise of social media platforms has significantly increased information accessibility while also facilitating the widespread dissemination of fake news, causing profoundly negative impacts on individuals and society. Current fake news detection methods are typically tailored for specific modality combinations (e.g., text-only, text-image, or text-video), making them inflexible in adapting to diverse modality combinations actually present on social media platforms. This limitation restricts the practical value of detection models in real-world scenarios. In this paper, we propose Uni-Detector, a universal multimodal fake news detection model. It customizes the Sparse Mixture-of-Experts (SMoE) layer as both the Feedforward Network (FFN) and Multi-head Self-Attention (MSA) modules within the Transformer architecture, tackling the aforementioned barriers. This approach separates the parameter space sufficiently for different modalities. The conditional computation nature of SMoE allows us to organically allocate relevant model components to dynamically accept and process any combination of input modalities. We conducted extensive experiments on Chinese and English multimodal datasets, each comprising three categories: text-only, text-image, and text-video. The proposed Uni-Detector achieves highly competitive performance across all modality combinations on test sets, establishing the Uni-Detector as a more flexible and powerful tool for detecting fake news in real-world scenarios.