Thinking Racial Bias in Fair Forgery Detection: Models, Datasets and Evaluations
Authors: Decheng Liu, Zongqi Wang, Chunlei Peng, Nannan Wang, Ruimin Hu, Xinbo Gao
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments conducted with 12 representative forgery detection models demonstrate the value of the proposed dataset and the reasonability of the designed fairness metrics. By applying the BPFA to the existing fairest detector, we achieve a new SOTA. Furthermore, we conduct more in-depth analyses to offer more insights to inspire researchers in the community. |
| Researcher Affiliation | Academia | Decheng Liu1*, Zongqi Wang2*, Chunlei Peng1 , Nannan Wang1, Ruimin Hu1, Xinbo Gao3 1Xidian University, Xi an, China 2Tsinghua University, Beijing, China 3Chongqing University of Posts and Telecommunications, Chongqing, China EMAIL, EMAIL, EMAIL, EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes methods like 'Bias Pruning with Fair Activations' and 'Approach Averaged Metric' using mathematical formulas and descriptive text, but no explicitly labeled 'Pseudocode' or 'Algorithm' blocks are present. |
| Open Source Code | No | The paper mentions reimplementing various forgery methods and applying them, but it does not include an explicit statement about releasing the source code for their own methodology (e.g., Fair FD dataset construction or BPFA algorithm) nor does it provide a link to a repository for their specific contributions. URLs provided are for third-party tools used. |
| Open Datasets | Yes | We use the RFW (Wang et al. 2019) dataset as pristine images. [...] Note that we still provide the original dataset with a resolution of 400 400 for scenarios requiring higher resolution. |
| Dataset Splits | Yes | Dataset. We use FF++ (c23) as our training set. Specifically, for each video, we select 32 frames, crop the facial region, and finally resize it to 256 256. We utilize the preprocessed data provided by (Yan et al. 2023b), which has already undergone the aforementioned operations. Our proposed new dataset serves as the testing set. |
| Hardware Specification | No | The paper describes experimental setup including dataset, algorithms, and training parameters, but does not provide specific details regarding the hardware (e.g., GPU models, CPU types, memory) used for conducting the experiments. |
| Software Dependencies | No | The paper mentions using the Adam optimization algorithm, but does not specify any software libraries, frameworks (like TensorFlow or PyTorch), or their corresponding version numbers used in the implementation. |
| Experiment Setup | Yes | In detail, these models are trained with the Adam optimization algorithm with a learning rate of 0.0002 and an epoch number of 10. The batch size is 32. And data augmentation methods including image compression, horizontal flip and rotation are applied. |