Article’s

The rapid increase of AI-generated text and manipulated images on social media platforms has raised serious concerns about the reliability of online information. Most existing fake news detection systems focus on a single modality, either text or images, which limits their ability to detect modern misinformation that often combines both. To address this, this paper proposes Fact Fusion, a multi-modal deep learning framework that integrates separate models for text and image analysis. For textual data, a pre-trained BERT-based encoder is used to generate contextual embeddings, which are further processed using a feed-forward neural network for classification. For visual data, a custom-designed Convolutional Neural Network (CNN) is employed to identify manipulated or fake images. Experimental results show that the text model achieves an accuracy of 91.3% with a weighted F1-score of 0.9256, while the image model achieves 93.8% accuracy with an F1-score of 0.9379. An additional comparison of embedding strategies demonstrates that mean pooling provides better performance than other methods. The complete system is deployed using a Streamlit-based interface, enabling real-time predictions with low latency. The proposed approach provides an efficient and scalable solution for detecting multi-modal misinformation in real-world scenarios.