【摘要】 随着短视频新闻的兴起，虚假短视频的检测成为亟待解决的问题。现有的短视频检测方法主要存在以下局限性:首先，传统方法多依赖于单一模态信息(如文字或视觉内容)，难以有效整合多模态特征，导致检测精度不足;其次，现有方法在处理短视频的复杂传播链和广泛传播影响范围时，缺乏对社会背景信息的深度挖掘，导致模型检测在这一领域关联度不够，难以应对如今造假手段高明的虚假短视频。本文基于深度学习和社会分析技术，自主优化形成最新的短视频数据集，设计了一种新的短视频虚假新闻检测架构SV-Shield系统。该系统将多尺度卷积与交叉注意力机制结合，有效地将社会背景的两个模态(如用户评论和传播路径)与其他模态(如视频内容和音频)进行关联。通过实验对比分析，验证了该系统相较于传统多模态检测方案，可以有效提升短视频虚假新闻的检测效率和精度。更多还原

【Abstract】 With the rise of short video news,the detection of fake short videos has become an urgent issue. Existing short video detection methods primarily suffer from the following limitations. First,traditional methods often rely on single-modal information(such as text or visual content),making it difficult to effectively integrate multimodal features and resulting in insufficient detection accuracy. Second,when handling the complex dissemination chains and extensive influence of short videos,current methods lack deep mining of social context information,which leads to inadequate relevance in this domain and makes it difficult to counter increasingly sophisticated fake short videos. Based on deep learning and social analysis techniques,a novel short video dataset is constructed through independent optimization,and a new architecture named the SV-Shield system for detecting fake news in short videos is proposed in this paper. The system integrates multi-scale convolution with a cross-attention mechanism,effectively correlating two social context modalities(such as user comments and propagation paths) with other modalities(such as video content and audio). Experimental comparative analysis verifies that,compared to traditional multimodal detection methods,the proposed system can effectively improve both the efficiency and accuracy of fake news detection in short videos.更多还原