Graphes hétérogènes de connaissances pour la détection des Fake News multimodales
Soutenue par Gaël Marec, à l’Université Paris 8, le 09/12/2025. Dirigée par Nédra Mellouli. Excellent sujet et d’actualité. Un spectre large sur les domaines abordés. Toujours un plaisir de travailler avec Nédra. [theses.fr] [Président du jury]
This doctoral dissertation presents an automatic approach for the detection of multimodal fake news, enriched by knowledge graphs. In the first part, we explain the context of the massification of information in which this thesis is situated. We present the main types of detection approaches (based on content, source, or context). In our state of the art, we present content-based approaches, focusing on textual and visual modalities, then those enriched by knowledge. We thus highlight the absence, in the literature, of a detection method combining multimodality, explainability, and integration of knowledge. …/…In the second part, we present the formal and applicative prerequisites of the model. We establish a method for detecting and localizing image transformations for the detection and explainability of the visual modality. We also propose tools based on the formalism of knowledge graphs and on the semantic understanding of large language models to structure information, whether visual or textual. We also explain the essential approaches for exploiting graphs through learned vector representations (e.g., TransE, R-GCN). …/…Finally, in the third part, we detail our complete detection system, which integrates a multimodal fusion module based on a simplified attention model. This module makes it possible to weight each signal (text, image, knowledge) and to produce a probability score that is explainable by the contributions of each modality.