Reproduce “DFGC 2021: A DeepFake Game Competition” 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
Bo Peng, Hongxing Fan, Wei Wang, Jing Dong, Yuezun Li, Siwei Lyu, Qi Li, Zhenan Sun and 15 more
The DeepFake Game Competition (DFGC) 2021, featuring teams from the Institute of Automation, Chinese Academy of Sciences and others, established an adversarial game between DeepFake creation and detection algorithms. This competition produced a challenging new dataset of DeepFakes designed to evade detection and demonstrated that current detection methods struggle to generalize against adaptively created fakes.
Key insights
- The median detection scores decreased over successive phases, indicating that creation methods became increasingly effective at evading detection.
- By the final phase, most fake datasets incorporated adversarial attacks or noise specifically designed to defeat detection methods.
- Models trained on the DFGC dataset showed poor generalization to other DeepFake datasets, highlighting the challenge of creating universally effective detection methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.