Research on detecting fake multimodal content on social media is growing in importance. Human-generated misinformation and AI-generated content generated by image synthesis models or visual-language models (VLMs) are the primary forms of deception. Existing research addresses these two types separately, limiting their effectiveness in real-world settings where specific types are unknown. To address this, this paper builds the OmniFake dataset, a comprehensive benchmark consisting of 127,000 samples that integrate human-curated misinformation collected from existing sources and newly synthesized AI-generated examples. Furthermore, we propose the Unified Multimodal Fake Content Detection (UMFDet) framework, designed to address both forms of deception. UMFDet leverages a VLM backbone augmented with a category-aware Mixture-of-Experts (MoE) Adapter and an Attribution Chain-of-Thought mechanism that provides implicit inference guidance to identify key deceptive signals. Experimental results demonstrate that UMFDet achieves robust and consistent performance across both types of misinformation, outperforming specialized baselines and providing a practical solution for real-world multimodal deception detection.
