Abstract:Multimodal media manipulation detection aims to model heterogeneous information such as images and text to achieve accurate authenticity assessment of image-text content. However, existing approaches have limited capability in modeling image frequency-domain features, making it difficult to fully capture fine-grained tampering traces. Meanwhile, they are insufficient at exploring logical forgery cues in textual content, such as internal contradictions and semantic inconsistencies, which restricts the effective use of discriminative information in the textual modality. To address these issues, this study proposes a multimodal media manipulation detection model, TruthSeeker, which consists of three core components: (1) a forgery clue mining module, which constructs a hierarchical feature learning mechanism comprising frequency-domain extraction, k-nearest neighbor (kNN)-based clustering, and feature compression, thus jointly modeling spatial- and frequency-domain information to enhance the perception of local tampering traces; (2) a textual contradiction enhancement module, which leverages a large language model to mine potential semantic conflicts from three aspects: factual inconsistency, emotional-context inconsistency, and colloquial expression, enhancing logical forgery cues in the textual modality; and (3) a deep forgery reasoning mechanism, which is based on cross-modal cross-attention to enable deep interaction between image and text forgery features, and performs collaborative optimization through a multi-task loss function. Experimental results demonstrate that TruthSeeker achieves an AUC of 99.49% in binary forgery detection and outperforms existing methods in multiple comparative experiments, validating the effectiveness of the model in multimodal media manipulation detection tasks.