Generating Attribution Reports for Manipulated Facial Images: A Dataset and Baseline

The project addresses a critical gap in image forensics: moving from binary 'fake/real' classification to explainable attribution. While most deepfake detection focuses on pixel-level artifacts, this project introduces a 'Why' component using natural language generation. Defensibility is low (3) because the project currently exists as an academic reference implementation and a dataset. While the FORGE dataset provides some data gravity, the methodology (combining localization with captioning) uses standard multimodal architectures that are easily replicable by well-funded labs. The 7 forks within 9 days of release indicate high academic interest, but the lack of stars suggests it hasn't yet crossed into broad developer utility. Frontier risk is medium because while OpenAI and Google are building general-purpose VLMs (GPT-4o, Gemini) that can reason about images, they often lack the specialized forensic training to detect high-end GAN/Diffusion manipulations. However, if these labs decide to prioritize 'safety and authenticity' features, this specific niche could be absorbed. Competitively, this sits adjacent to specialized deepfake detection startups like RealityDefender or Sentinel, but focuses more on the *explanation* layer which is vital for legal and journalistic contexts. Its primary threat is the rapid advancement of general-purpose visual reasoning models which may eventually achieve similar attribution capabilities without specialized forensic datasets.