AI-Face: A Million-Scale Demographically Annotated AI-Generated Face Dataset and Fairness Benchmark

TL;DR

The paper tackles biased detection of AI-generated faces by introducing AI-Face, a million-scale dataset with demographically annotated real and AI-generated faces across Deepfake Videos, GANs, and Diffusion Models, enabling a comprehensive fairness analysis. It trains a CLIP-based lightweight annotator with an imbalance loss $L_{imb}$ and a Sinkhorn-based fairness loss $L_{fair}$, and improves generalization via Sharpness-Aware Minimization (SAM). Based on AI-Face, it launches a fairness benchmark evaluating 12 detectors across four model types using five fairness and five utility metrics, revealing that fairness-enhanced methods offer gains but biases persist across generative methods and post-processing can degrade utility. The work highlights the need for robust, generalizable fair detectors and suggests integrating foundation models as a promising direction to mitigate bias and improve cross-domain fairness in AI-face detection.

Abstract

AI-generated faces have enriched human life, such as entertainment, education, and art. However, they also pose misuse risks. Therefore, detecting AI-generated faces becomes crucial, yet current detectors show biased performance across different demographic groups. Mitigating biases can be done by designing algorithmic fairness methods, which usually require demographically annotated face datasets for model training. However, no existing dataset encompasses both demographic attributes and diverse generative methods simultaneously, which hinders the development of fair detectors for AI-generated faces. In this work, we introduce the AI-Face dataset, the first million-scale demographically annotated AI-generated face image dataset, including real faces, faces from deepfake videos, and faces generated by Generative Adversarial Networks and Diffusion Models. Based on this dataset, we conduct the first comprehensive fairness benchmark to assess various AI face detectors and provide valuable insights and findings to promote the future fair design of AI face detectors. Our AI-Face dataset and benchmark code are publicly available at https://github.com/Purdue-M2/AI-Face-FairnessBench

Figures (20)

• Figure 1: Comparison between AI-Face and other datasets in terms of demographic annotation, generation category, and the number of generation methods. 'DF', 'GAN', and 'DM' stand for Deepfake Videos, Generative Adversarial Networks, and Diffusion Models.
• Figure 2: Generation pipeline of our Demographically Annotated AI-Face Dataset. First, we collect and filter face images from Deepfake Videos, GAN-generated faces, and DM-generated faces found in public datasets. Second, we perform skin tone, gender, and age annotation generation. Skin tone is estimated by combining facial landmark detection with color analysis to generate the corresponding annotation. For gender and age, we develop annotators trained on the IMDB-WIKI dataset rothe2015dex, then use them to predict attributes for each image.
• Figure 3: Distribution of face images of the AI-Face dataset. The figure shows the (a) subset distribution and the demographic distribution for (b) skin tone, (c) gender, and (d) gender. The outer rings in (b), (c), and (d) represent the proportion of groups within each attribute category, while the inner rings indicate the distribution of fake (F) and real (R) images within those groups.
• Figure 4: Visualization of the intersectional $F_{EO}$ (%) and AUC (%) of detectors on different subsets. The smaller $F_{EO}$ polygon area represents better fairness. The larger AUC area means better utility.
• Figure 5: FPR(%) of each intersectional subgroup The dashline represents the lowest FPR on Female-Light (F-L) subgroup.