Competition

News

2022.6.13: Top-3 teams are announce in the "Evaluation" page here.

2022.4.26: Important notices on the final phases here.

2022.4.26: The 3rd update of the testing data is available.

2022.4.11: The 2nd update of the testing data is available.

2022.3.29: The 1st update of the testing data is available.

Overview

DeepFake is developing fast, and realistic face-swaps are growingly deceiving and hard to be detected. On the contrary, DeepFake detection methods are also improving. There is a two-party game between DeepFake creators and defenders. We are organizing this competition to provide a common platform for benchmarking the game between current state-of-the-art DeepFake creation and detection methods. The main research question to be answered by this competition is the current status of the two adversaries against each other. This is the second edition after the DFGC-21 held in the last year, with many improvements. By organizing this competition, we hope to further stimulate research ideas to build better defenses against DeepFake threats.

This is the DeepFake detection track of DFGC-22. For the creation track, please see here.

Dataset

The test dataset for the DeepFake detection track will be drawn from our collected material videos (as real samples) and from the face swap videos created both by the creation track participants and by the organizers (as fake samples). These videos may undergo random post-processing operations to imitate real-world video quality degradation during transmission. For the final testing stage, additional real and fake videos from other sources may also be included for more data diversity.

After registration to this track, you can download the dataset from "Participate >> Get Data"

Competition Protocols

The competition is composed of two adversarial tracks: DeepFake creation and DeepFake detection, which are conducted in parallel.

For the Deepfake detection track, participants are required to develop their detection methods and models using any publicly available datasets, except the provided competition datasets. The data collected and created in this competition CANNOT be used for training or developing detection models, and they are only for testing. There are two stages in the detection track. The first one is the validation stage, where the models are tested on the public test set. The second one is the final stage, where the models are tested on the private test set. The public and the private test sets have disjointed IDs, they absorb submissions from the creation track and will be updated after each evaluation round of the creation track. This process makes our competition a dynamic game between the two tracks.

During the validation stage, the public test set, which is composed of real and fake video clips without groundtruth labels, will be released to the detection track participants for producing inference results. The inference results, in the form of text format indicating the predicted fake probability of each video clip, need to be submitted to the competition platform for evaluation, and the leaderboard (LB) will be automatically updated with each team's best score. After the validation stage ends, the top-10 teams on the LB will enter the final stage, where these models will be tested on the private set to obtain the final score. The final stage models need to be contained in a docker and sent to the organizers to complete the final stage testing. Potential winning teams need to send their training code and training data, accompanied with a technical report, to the organizers to check the reproducibility and the compliance with the competition rules.

Timeline

3.15 - 5.4 5.7:  The validation phase. Up to 2 submissions can be made in each day.

5.5 5.8- 5.15: The final phase. Top-10 teams on the LB send their LB models to the organizers.

Awards

Top-3 winners in the final phase will receive award certificates and a bonus.
A total of 50,000 CNY (around 8,000 USD) bonus will be rewarded to winners in the two tracks of DFGC-22.
The Top-3 of this track will receive: 12,000 CNY, 8,000 CNY, 5,000 CNY

Organizing Committee

Chair
Prof. Jing Dong, IEEE Senior Member, from CASIA

Members
Dr. Bo Peng, from CASIA
Prof. Wei Wang, from CASIA
Prof. Zhen Lei, IAPR Young Biometrics Investigator Awardee, from CASIA
Prof. Zhenan Sun, IAPR Fellow, from CASIA
Prof. Siwei Lyu, IEEE Fellow, from University at Buffalo, SUNY

Sponsor

Tianjin Academy for Intelligent Recognition Technologies