Reproduce “All in One: Unifying Deepfake Detection, Tampering Localization, and Source Tracing with a Robust Landmark-Identity Watermark” 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
Junjiang Wu, Liejun Wang, Zhiqing Guo
Researchers at Xinjiang University introduce LIDMark, a unified proactive forensics framework that simultaneously performs deepfake detection, tampering localization, and source tracing using a robust landmark-identity watermark. This framework achieves high visual quality with PSNR 44.31 and SSIM 0.99, while enabling robust source tracing with an average Bit Error Rate of 2.55% against deepfake manipulations and effective tampering localization.
Key insights
- A single, composite watermark can effectively carry heterogeneous forensic information (tamper-sensitive geometric landmarks and a robust source identifier) while maintaining high imperceptibility and robustness against diverse manipulations.
- A factorized decoding architecture, leveraging a shared feature backbone with task-specific heads, efficiently allows for the synchronous and robust recovery of mathematically distinct watermark components (regression for landmarks, classification for identifiers).
- Comparing robustly recovered 'intrinsic' embedded geometric information with 'extrinsic' re-detected features provides a powerful blind-reference mechanism for both global deepfake detection and fine-grained tampering localization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.