This competition will release the Multi-dimensional Facial Forgery Dataset (MultiFF), which specifically includes the multi-dimensional Facial Forgery Image Dataset (MultiFFI) and the Multi-dimensional Facial Forgery Video Dataset (MultiFFV). These two subsets are used for the image Deepfake detection task in Track 1 and the audio-video Deepfake detection task in Track 2, respectively.
The MultiFF data set is built based on the "Technical Specifications for Financial Applications of Fake Digital Face Detection" launched by Ant Group. At the same time, it is deeply combined with the actual defensive experience in the digital business to include truly diversified deep forgery attacks, which mainly contain:
1. Wide range of attack types: including face-swapping, activation, attribute editing, full-face synthesis, audio-driving face generation, face restoration, digital adversarial samples, automated PS, and other common types;
2. Multiple generation methods: There are more than 50 Deepfake generation methods applied to the entire data set. In addition to the classic GANs model, the generation model also fully covers the new generative AI model of the Diffusion series. The attack generation paradigm includes image-to-image generation, text-to-image generation, text-to-speech generation, text-to-video generation, etc.
The test set will cover new Deepfake generation methods to simulate the emerging new attack methods in real scenarios and test the generalization ability and practical application performance of the defense algorithm. In addition, in order to ensure the diversity of samples, we have established a test set including race, age, expression, light, occlusion, indoor and outdoor scenes, and other factors to fully test the effectiveness of the detection methods in various scenes and different demographic groups.
Different from the existing open source Deepfake data sets such as FaceForensics++, DiFF, LAV-DF, and AV-Deepfake1M, which only focus on a certain type of generation model or a certain part of the generation method, the MultiFF data set released in this competition is much more comprehensive and diversified in multiple dimensions such as types, generation methods, generation paradigms, and face material attributes.

**Track 1: Deepfake Image Detection**
The task of the competition is to determine whether a face image is a Deepfake image and output its probability score of being a Deepfake image. 

The competition is divided into three phases:
Phase 1: The training set and validation set are released. The training set train_label.txt is used to train the model, and the validation set val_label.txt is only used for model tuning. For example, each line in train_label.txt or val_label.txt contains two parts separated by commas. The first part is the image file name, and the second part is the label value (label=1 means a Deepfake image, label=0 means a real face image).


train_label.txt
```js
00000001.jpg,1  # the fisrt line of train_label.txt
00000002.jpg,0  # the second line of train_label.txt
…
```
val_label.txt
```js
00000003.jpg,1  # the fisrt line of val_label.txt
00000004.jpg,0  # the second line of val_label.txt
…
```

Phase 2: After Phase 1, we will release the test set in Phase 2. Contestants need to submit the prediction score file prediction.txt of the test set in the system, and we will feedback on the test score results online. Each line in the file contains two parts separated by commas. The first part is the image file name, and the second part is the Deepfake score corresponding to the model prediction (i.e. the probability value that the sample belongs to a Deepfake image).
prediction.txt
```js
00000005.jpg,0.123456
00000006.jpg,0.123456
…
```

Phase 3: After Phase 2, the top 10 teams can advance to the final Phase 3. At this stage, contestants need to submit code docker and technical reports. Docker requirements include the original training code and test API (the function input is the image path, and the output is the deepfake score predicted by the model). We will check and rerun the algorithm code to reproduce the training process and test results. The final ranking will be based on the test set scores submitted by the players, the test set scores we reproduced, and the weighted scores of the technical reports, with weights of 0.4, 0.4, and 0.2 respectively.

**Track 2: Deepfake Audio-Video Detection**
Coming soon