摘要： 网络谣言的传播形式已从单一文本演化为包含图像、视频等多模态信息的复合形态。多模态内容凭借更强的视觉冲击力和情感煽动性，显著提升了虚假信息的传播速度与覆盖范围，给网络空间治理带来了严峻挑战。本文提出一种基于特征增强与双路决策融合的检测模型。模型整体架构主要由轻量化视觉特征增强、双路协同判别以及自适应决策融合组成。首先，设计轻量化平行多尺度特征增强模块（Lightweight Parallel Multi-scale Feature Enhancer, LPMFE），通过全局自注意力、空间深度可分离卷积与时序一维卷积三条分支分别捕获不同维度特征，并利用可学习权重动态融合，配合门控残差机制与前馈网络，在保留主干特征完整性的同时实现图像局部特征的精细化增强。然后，构建双路协同判别机制：生成式分支首先通过基于交叉注意力的图像投影器，将视觉特征序列对齐至大语言模型的词嵌入空间，而后将投影后的视觉Token插入文本序列，输入大语言模型进行自回归推理，提取最后一个Token的隐藏状态经线性分类头映射为词表逻辑值，最终通过Softmax函数转化为生成式先验概率，以利用模型参数化知识进行宏观逻辑推断。判别式分支为显式校验图文语义一致性，共享同一大语言模型深层的隐藏状态，并通过掩码聚合得到全局文本向量；同时将增强后的视觉特征进行池化，二者经投影网络映射至256维共享语义空间；在此空间中引入监督对比损失以优化模态特征的可区分性，进而计算投影特征的绝对差异与元素级乘积以捕捉细粒度语义冲突，将原始投影特征与两类交互特征在通道维度拼接后输入多层感知机，输出图文一致性判别分数。最后，在决策融合层面，引入可学习的权重参数对双路输出概率进行自适应融合，并设置双阈值构成三区间判定逻辑，将样本划分为高置信度谣言、高置信度非谣言及模糊待定三类；同时在优化目标中引入边界惩罚项以防止阈值坍塌，确保融合策略的稳定收敛。在Weibo、Fakeddit和FineFake数据集上进行实验。Weibo数据集上，准确率达93.23%，非谣言精准率/召回率为95.72%/90.50%，谣言为90.99%/95.95%。Fakeddit上准确率91.56%，非谣言与谣言F1值分别为92.98%和89.42%，差值3.56个百分点。FineFake上准确率81.21%，非谣言/谣言F1为83.42%/78.31%。消融实验表明：移除LPMFE后FineFake准确率降至79.20%，谣言召回率下降5.32个百分点；将交叉注意力投影器替换为线性映射后Fakeddit谣言F1下降6.01个百分点；移除判别分支后Weibo准确率降至86.59%，谣言召回率下降16.27个百分点。本研究表明，通过在视觉端引入特征增强机制来抑制信息冗余与复杂场景干扰，并结合大语言模型的参数化常识推理与异构模态间显性一致性校验构建生成式与判别式双路协同判别机制，能够克服传统单范式检测模型的局限性，有效纠正大语言模型的幻觉与单模态语义偏置，显著提升多模态谣言检测的综合精度。

Abstract: The propagation of online rumors has evolved from single-modal text to composite forms incorporating multimodal information such as images and videos. Leveraging stronger visual impact and emotional appeal, multimodal content significantly accelerates the dissemination speed and expands the coverage of false information, posing severe challenges to cyberspace governance. This paper proposes a detection model based on feature enhancement and dual-path decision fusion. The overall architecture consists of three core modules: lightweight visual feature enhancement, dual-path collaborative discrimination, and adaptive decision fusion. First, a Lightweight Parallel Multi-scale Feature Enhancer (LPMFE) is designed, which employs three parallel branches—global self-attention, spatial depthwise separable convolution, and temporal 1D convolution—to capture features from different dimensions. A learnable weight vector dynamically fuses these multi-scale outputs, while a gated residual mechanism and feedforward network refine local image features while preserving the integrity of the backbone representation. Second, a dual-path collaborative discrimination mechanism is constructed. The generative path first aligns the visual feature sequence into the word embedding space of a large language model (LLM) via a cross-attention based image projector. The projected visual tokens are then inserted into the text sequence and fed into the LLM for autoregressive inference. The hidden state of the last token is extracted and mapped through a linear classification head to logits over the vocabulary, which are finally converted into a generative prior probability via a Softmax function, leveraging the model’s parametric knowledge for macroscopic logical reasoning. The discriminative path explicitly verifies image–text semantic consistency by sharing the deep hidden states from the same LLM and aggregating them via masking to obtain a global text vector; meanwhile, the enhanced visual features are pooled. Both modalities are projected into a shared 256-dimensional semantic space. A supervised contrastive loss is introduced to optimize the discriminability of modal features. The absolute difference and element-wise product of the projected features are computed to capture fine-grained semantic conflicts. The original projected features and the two interaction features are concatenated along the channel dimension and fed into a multilayer perceptron to output a consistency discrimination score. Finally, at the decision fusion level, a learnable weight parameter is introduced to adaptively fuse the dual-path output probabilities. A two-threshold mechanism is set to form a three-interval decision logic, categorizing samples into high-confidence rumors, high-confidence non-rumors, and ambiguous cases. A boundary penalty term is added to the optimization objective to prevent threshold collapse and ensure stable convergence of the fusion strategy. Experiments are conducted on the Weibo, Fakeddit, and FineFake datasets. On Weibo, the method achieves an accuracy of 93.23%, with precision/recall for non-rumors at 95.72%/90.50% and for rumors at 90.99%/95.95%. On Fakeddit, the accuracy reaches 91.56%, with F1-scores of 92.98% for non-rumor and 89.42% for rumor, a difference of 3.56 percentage points. On FineFake, the accuracy is 81.21%, with non-rumor/rumor F1-scores of 83.42%/78.31%. Ablation studies demonstrate that removing LPMFE decreases FineFake accuracy to 79.20% and rumor recall by 5.32 percentage points; replacing the cross-attention projector with a linear mapping reduces Fakeddit rumor F1 by 6.01 points; and removing the discriminative branch lowers Weibo accuracy to 86.59% and rumor recall by 16.27 points. This study demonstrates that incorporating a visual feature enhancement mechanism to suppress information redundancy and complex scene interference, combined with an LLM-based parametric commonsense reasoning and explicit cross-modal consistency verification via a generative-discriminative dual-path collaborative design, overcomes the limitations of single-paradigm detection models, effectively mitigates LLM hallucinations and unimodal semantic bias, and substantially improves the overall accuracy of multimodal rumor detection.