Existing fake news detection methods suffer from notable limitations. They typically fail to adequately explore multimodal correlations, overlook critical information, thereby introducing redundancy, and disregard the positive or negative influence of similar news articles, all of which impair detection accuracy. Moreover, approaches capable of synergistically utilizing textual, visual, and social-graph modalities remain scarce. To address these issues, this study proposes a multimodal fake-news detection model integrating social graphs and label optimization. The method simultaneously extracts features from text, images, and social graphs. Key social graph signals are enhanced via a global dynamic weighting mechanism, while visual representations are strengthened using image descriptions. Cross-modal complementary advantages are achieved through cross-modal cross-attention, and contrastive learning is applied for cross-modal alignment. In the label optimizer, bidirectional label attention is used to capture positive and negative correlations among semantically similar news articles, thereby refining the predicted labels. Experimental results on Weibo and PHEME show that the proposed model improves accuracy by 1.63 % and 3.01 %, and F1-score by 2.46 % and 3.67 %, respectively, compared to baseline methods.