摘要
Abstract
[Purpose]This paper aims to construct an identification method for multimodal disinformation tailored to the application require-ments of cognitive warfare.[Method]A multimodal causal inference-based method for recognizing disinformation in cognitive warfare(MCI-CWDI)was proposed.Firstly,targeted fine-tuning of the CLIP large model was performed to extract multimodal causal consisten-cy features.Secondly,the PC algorithm was adopted to construct a structural causal model(SCM)incorporating latent variables of objec-tive real events,thereby realizing the preliminary recognition of disinformation.Finally,a multi-dimensional causal contradiction detec-tion algorithm covering temporal,spatial,and logical aspects was designed to further enhance the disinformation recognition performance.[Result/Conclusion]The results showed that the F1-score of MCI-CWDI on the Fakeddit test set outperformed the optimal baseline BLIP-FT by 1.5 percentage points,and its F1-score on the cross-lingual generalization test set of MM-COVID was higher than that of the sin-gle-modal causal model Causal-BERT.Ablation experiments indicated that the causal intervention module,contradiction detection mod-ule,CLIP fine-tuning module,and latent variable modeling all had a facilitating effect on the model performance,with differences in their contribution degrees.Further research revealed that the model exhibited excellent robustness in scenarios involving text disturbance,image disturbance,and multi-modal joint disturbance.关键词
认知战/多模态虚假情报/因果推断/MCI-CWDI/结构因果模型/跨域泛化Key words
cognitive warfare/multimodal disinformation/causal inference/MCI-CWDI/structural causal model/cross-domain gener-alization分类
信息技术与安全科学
