ScholarWorks@성균관대학교

초록

The rapid and uncontrolled spread of deepfake content on social media poses growing risks to public trust, personal safety, and digital information ecosystems. Despite extensive research on post-hoc detection models, proactive methods that anticipate both the scale and structural patterns of manipulated media diffusion remain largely underexplored. In this work, we introduce a multimodal spatiotemporal model that forecasts deepfake propagation using early sharing activity and the perceptual characteristics of the manipulated content. The proposed model integrates a Temporal Graph Transformer to encode the evolving structure of early propagation cascades, a FiLM-based visual encoder that modulates image representations on demographic and perceptual metadata, and two contrastive self-supervised objectives that capture propagation dynamics and cross-modal consistency. The forecasting task is formulated as a multi-task problem, jointly predicting the final cascade size and the Wiener Index, a graph-theoretic measure of structural diffusion complexity. To validate our model, we construct a real-world dataset of deepfake content and its propagation traces on a social media platform. Experiments demonstrate that our approach consistently outperforms competitive baselines. This work takes a step toward early intervention systems that mitigate the societal harms associated with rapidly spreading synthetic media.

키워드