A bi-level multi-modal fake generative news detection approach: from the perspective of emotional manipulation purpose

As conversational bot Large Models become a daily channel available to everyone, fake Artificial Intelligence Generated Contents (fake AIGCs) have emerged as a serious threat in cyberspace security, with severity varying significantly across regional and cultural contexts, making upgrading detectors for this novel scenario an urgent need. Compared to manually fabricated fake news, fake AIGCs tend to be modality-rich and logically complete, which poses a challenge to traditional detection based on cross-modal semantic consistency. Unlike existing works that mostly focus on analyzing what is generated, we introduce a new perspective that considers the fake AIGCs why generated. By aiming at the “emotional evoking and manipulation” behind fake AIGCs, we design EmoDect, a bi-level (posts and comments) consistency integration model for fake AIGCs detection, which firstly introduces Large Language Models (LLM) for comments generation to imitate the public emotions specific to each modality while eliminating the regional and national differences, and considers the emotional consistency via (1) inner-level posts consistency cross modalities; (2) inner-level comments consistency cross modalities; (3) intra-level consistency within each modality; and (4) intra-level consistency of (1) and (2) aggregations. The proposed EmoDect outperforms baseline models in two real-world datasets by 2.48% and 1.27% in terms of accuracy, highlighting the significance of bi-level emotional capabilities in improving fake AIGCs detection.