Reproduce “A Sanity Check for Multi-In-Domain Face Forgery Detection in the Real World” 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

Jikang Cheng, Renye Yan, Zhiyuan Yan, Yaozhong Gan, Xueyi Zhang, Zhongyuan Wang, Wei Peng, Ling Liang

This paper introduces Multi-In-Domain Face Forgery Detection (MID-FFD), a paradigm that focuses on developing robust detectors for extensive, diverse known forgery types rather than relying on idealistic generalization. The proposed DevDet framework enhances real/fake distinction over domain differences, leading to improved absolute classification accuracy on mixed deepfake datasets while preserving generalization to unseen forgeries.

Key insights

  • The 'domain-dominant issue' is a critical barrier to real-world deepfake detection, as it causes detectors to prioritize inter-domain variations over true real/fake differences in the feature space.
  • A targeted 'developing' process, akin to photographic development, can effectively make subtle deepfake traces more pronounced in images, aiding detector discrimination.
  • An adaptive dose mechanism, informed by a dictionary of hard-fake characteristics, is essential for simultaneously boosting performance on diverse in-domain data and preserving generalization to unseen deepfake types.

Summary by alphaXiv.

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