ETD

• computer vision
• fake news
• image manipulation
• multimodal deepfake detection

This thesis focuses on deepfake detection within the specific context of fake news, an area of increasing concern in digital media integrity. It firstly enhances the existing Fakeddit fake news dataset by incorporating synthetically generated images, thereby creating a more challenging and comprehensive benchmark for detection algorithms. A comprehensive comparative analysis was then conducted to evaluate the effectiveness of unimodal (image-only) and multimodal (image+text) models in detecting deepfakes. The study compares different architectural frameworks, specifically one based on CLIP and another leveraging ResNet and BERT, to determine which was most effective in this context. Furthermore, the research demonstrates that deepfakes paired with fake news text pose a greater detection challenge compared to those associated with truthful text.