The DFGC 2022 was co-organized by CRIPAC, CASIA and SUNY Buffaloand. It was accepted by the 2022 International Joint Conference on Biometrics (IJCB), which is the premier conference in the biometrics research field. The competition was recently successfully held (https://codalab.lisn.upsaclay.fr/competitions/3523), and it drew attention from more than 50 registered teams all over the world.

The DeepFake is rapidly evolving, and realistic face-swaps are becoming more deceptive and difficult to detect. On the other hand, methods for detecting DeepFakes are also improving. There is a two-party game between DeepFake creators and defenders. This competition provides a common platform for benchmarking the game between the current state-of-the-arts in DeepFake creation and detection methods. The main research question to be answered by this competition is the current state of the two adversaries when competed with each other. This is the second edition after the last year's DFGC 2021, with a new, more diverse video dataset, a more realistic game setting, and more reasonable evaluation metrics. With this competition, we aim to stimulate research ideas for building better defenses against the DeepFake threats. We also release our DFGC 2022 dataset contributed by both our participants and ourselves to enrich the DeepFake data resources for the research community (https://github.com/NiCE-X/DFGC-2022).

The competition had two tracks: the DeepFake Creation (DC) track and the DeepFake Detection (DD) track. The DC track was composed of three submission rounds, and the DD track was composed of the validation phase and the final phase. The two sides were evaluated against each other. We built a face-swap DeepFake dataset containing 4394 video clips, together with our participants. This dataset features diverse face-swap methods and post-processing operations, and we release this dataset for future DeepFake related research.

Table 1. The statistics of the new DFGC 2022 deepfake dataset

We received 35 submissions in the creation track and 25 submissions in the detection track. The top-3 teams in both tracks were awarded with bonus sponsored by the Tianjin Academy for Intelligent Recognition Technology. Summarizing the top solutions, it was found that top detection solutions all used ensembles of most recent effective deep models, e.g. vision transformers. They also trained the models on many deepfake datasets and used various data augmentations. In the creation track, various state-of-the-arts face-swap methods were adopted by the creation participants, and careful post-processing operations were adopted to either improve the deception ability to humans (e.g. de-blurring, super-resolution, advanced blending or merging methods) or to detection models (e.g. compression). Comparing the two sides, it was shown that the best detection models still struggle when facing the best quality deepfake videos and suffer from generalization problems.

Table 2. Scores of the top-3 creation track teams

Table 3. Scores of the top-3 detection track teams on different test sets

The competition summary paper has been accepted by IJCB 2022, and the preprint can be found here at https://arxiv.org/abs/2206.15138.