Reproduce “DiffusionFake: Enhancing Generalization in Deepfake Detection via Guided Stable Diffusion” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ke Sun, Shen Chen, Taiping Yao, Hong Liu, Xiaoshuai Sun, Shouhong Ding, Rongrong Ji

DiffusionFake enhances deepfake detection generalization by inverting the deepfake generative process using a pre-trained Stable Diffusion model, guiding the detector to learn disentangled source and target identity features. This plug-and-play framework significantly improves cross-dataset and cross-manipulation performance, boosting average AUC by approximately 6% on various unseen deepfake datasets without adding inference overhead.

Key insights

  • Deepfake images inherently contain amalgamated information from both source and target identities, a universal characteristic that can be leveraged for detection regardless of the specific manipulation technique.
  • By inverting the deepfake generative process and compelling a detection network to facilitate the reconstruction of original source and target images, more robust and generalizable features can be learned.
  • Pre-trained diffusion models can effectively serve as a powerful 'guide' during training to disentangle complex image features, enhancing the detector's ability to discriminate between real and fake content.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.