Reproduce “DF-Net: The Digital Forensics Network for Image Forgery Detection” 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
David Fischinger, Martin Boyer
DF-Net, developed by the Austrian Institute of Technology, presents a deep learning model for pixel-wise image forgery detection, demonstrating superior accuracy and unprecedented robustness against image degradation from social network processing. The model also achieves a significant reduction in processing time compared to prior state-of-the-art methods.
Key insights
- Integrating scSE blocks and increasing input image size to 256x256 pixels substantially enhances detection performance, improving mean AUC by 0.082 and 0.029, respectively.
- Pre-scaling images to 256x256 pixels during training is a crucial factor for DF-Net's exceptional robustness against lossy OSN operations, preserving manipulation traces.
- Combining specialized sub-network predictions using a maximum-operator strategy (max(M1,M2)) is more effective than averaging, achieving superior overall accuracy by capturing complementary forgery evidence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.