🎭 Master Thesis - Combining Deepfake Detection with Fake News Analysis

Thesis Title:

On the Effectiveness of Deepfake Detection on Multimodal Fake News

My Master Thesis, conducted in collaboration with ISTI-CNR of Pisa, focused on the topic of Deepfake Detection. Today, the misinformation spreads rapidly, detecting deepfakes and fake news has become crucial. However, there was a lack of research on combining these two areas. My thesis aims to bridge this gap by proposing a new type of analysis that combines the detection of deepfakes and fake news.

Objectives: 🎯

The primary aims of my thesis were to:

• Propose an innovative analysis about the effectiveness of deepfake detection in the context of fake news.
• Create a brand new dataset that integrates both deepfake images and fake news text, named DeepFakeNews.
• Develop a multimodal deepfake detector, robust towards fake news context.

Contributions: 🔥

• Innovative Analysis: Proposed an innovative analysis to evaluate the effectiveness of deepfake detection tools when used in conjunction with fake news.
• DeepFakeNews Dataset: Introduced a new public dataset designed to support the detection of both deepfakes and fake news. This dataset, is an expansion of a pre-existing fake news detection dataset called Fakeddit.

DeepFakeNews is publicly available on Zenodo (HERE) and includes:

• Three CSV files for training, testing, and validation.
• Zip files containing image splits.

• 509,916 images, balanced between 254,958 authentic and 254,958 deepfake images generated via Stable Diffusion 2, Dreamlike, and GLIDE models.

• Multimodal Deepfake Detection Model: Proposed a robust multimodal detection model combining ResNet50 for image analysis and BERT for text analysis, achieving an impressive accuracy of 93.3%.

Dataset Overview: 📊

The DeepFakeNews dataset was made to address modern misinformation challenges, that include both visual and textual deception.

Applications: 🔍

This dataset is ideal for:

• Standard deepfake detection.
• Fake news detection.
• Combined multimodal analyses of visual and textual content, providing a robust framework for evaluating detection systems.

Impact: 📈

Since its release, the DeepFakeNews dataset has been downloaded nearly 100 times, serving as a crucial benchmark for ongoing research in digital misinformation detection.

🔗 GitHub Repository

Visit the thesis repository here for accessing the codebase and models.