Students

This project addresses the urgent need for a modern, representative dataset to advance deepfake detection research. As AI-generated images from proprietary models like Midjourney, Ideogram, and FLUX.1 pro have reached unprecedented realism, traditional detection methods—developed using outdated datasets from open-source models—are becoming ineffective. This gap hinders the development of robust detection algorithms and raises questions about the feasibility of large-scale deepfake detection, especially given the implications for misinformation and election security.

The primary goal of the project is to create a comprehensive benchmark dataset featuring AI-generated images from both leading proprietary and open-source models. The dataset will be diverse, including human portraits, landscapes, and objects, with a particular focus on images most likely to be used for misinformation. Each image will be systematically labeled by generation settings, model, and content type, enabling detailed analysis.

Key objectives include:
• Collecting a broad and current set of AI-generated images that reflect state-of-the-art generative techniques.
• Evaluating whether images from advanced proprietary models exhibit detectable artifacts.
• Assessing the effectiveness and generalizability of existing deepfake detection methods on this new dataset.
• Investigating which types of images and content present the greatest challenges for detection.
• Providing insights to guide the development of more effective detection strategies, particularly for high-stakes contexts like elections.

By filling this critical gap, the project will equip researchers and practitioners with the tools needed to accurately evaluate and improve deepfake detection, supporting efforts to safeguard public trust and media integrity in the age of advanced AI imagery.