Confidence Breeds Success: Improving Fake News Video Detection via LVLM-Assisted Inference

The rapid rise of short video platforms worldwide has brought the risk of widespread dissemination of fake news. Current approaches typically employ fine-tuned small models to detect news videos, which have significant limitations. Utilizing the knowledge of LLMs has been empirically proven to be a promising direction, but it faces constraints and hallucination problems. To address these issues, this paper proposes Improving Fake News Video Detection via LVLM-Assisted Inference (IFAI). Specifically, this paper introduces a news video semantic understanding approach to generate auxiliary information. Then, key information selection and learning modules are designed to bridge the gap between LVLMs and small models, improving the efficiency of utilizing the supplementary knowledge of LVLMs. To the best of our knowledge, this is the first paper to explore the application of LVLMs in fake news video detection. Extensive experiments demonstrate that IFAI achieves state-of-the-art performance on the benchmark datasets.