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

Problem Definition. There are existing online software (e.g., Face++ [42]) and open-source tools (e.g., InsightFace [43]) for face attribute prediction. However, they fall short of our task due to two reasons: 1) They are mostly designed for face recognition and trained on datasets of real face images but lack generalization capability for annotating AI-generated face images. 2) They do not provide uncertainty scores for their predictions that can be used to identify mispredicted samples for further annotation correction. Given a training dataset $\mathbb{D}=\{(X_{i},G_{i},A_{i},R_{i})\}_{i=1}^{n}$ with size $n$, where $X_{i}$, $G_{i}\in\{Female,Male\}$, $A_{i}\in\{Young,Middle\text{-}aged,Senior,Others\}$, and $R_{i}\in\{Asian,White,Black,Others\}$ represent the $i$-th face image, and its gender, age, and race labels/attributes, respectively. Our goal is to design a lightweight, generalizable annotator based on $\mathbb{D}$ to predict facial demographic attributes with uncertainty scores for each face image in our dataset.

Annotator. Architecture: We utilize CLIP [27] for its strong zero-shot and few-shot learning capabilities. Leveraging CLIP’s pre-training on diverse datasets, we create a lightweight annotator for facial images. Our annotator employs a frozen pre-trained CLIP ViT L/14 [44] as a feature extractor $\mathbf{E}$ followed by a trainable 3-layer Multilayer Perceptron (MLP) as a multi-task (i.e., gender, age, and race prediction) classifier parameterized by $\theta$. Loss: For each image $X_{i}$, its feature $f_{i}$ is obtained through $f_{i}=\mathbf{E}(X_{i})$ and then is fed into the MLP multi-task classifier with conventional classification losses for face attribute prediction. The learning objective is formulated as: $\mathcal{L}(\theta)=C(\widetilde{h}(f_{i}),G_{i})+C(\overline{h}(f_{i}),A_{i})+C(\widehat{h}(f_{i}),R_{i})$, where $C(\cdot,\cdot)$ represents the (binary) cross-entropy (CE) loss. $\widetilde{h}$, $\overline{h}$, and $\widehat{h}$ represent the classification heads for gender, age, and race, respectively. Optimization: Traditional optimization methods like stochastic gradient descent can lead to poor model generalization due to sharp loss landscapes with multiple local and global minima. To address this, we use Sharpness-Aware Minimization (SAM) [45] to enhance our annotator’s generalization by flattening the loss landscape. Specifically, flattening is attained by determining the optimal $\epsilon^{*}$ for perturbing model parameters $\theta$ to maximize the loss, formulated as: $\epsilon^{*}=\arg\max_{\|\epsilon\|_{2}\leq\gamma}{\mathcal{L}}\textbf{(}\theta+\epsilon\textbf{)}\approx\arg\max_{\|\epsilon\|_{2}\leq\gamma}\epsilon^{\top}\nabla_{\theta}\mathcal{L}=\gamma\texttt{sign}(\nabla_{\theta}\mathcal{L})$, where $\gamma$ controls the perturbation magnitude. This is approximated using a first-order Taylor expansion, assuming $\epsilon$ is small. The final equation is obtained by solving a dual norm problem, where sign represents a sign function and $\nabla_{\theta}\mathcal{L}$ being the gradient of $\mathcal{L}$ with respect to $\theta$. As a result, the model parameters are updated by solving: $\min_{\theta}\mathcal{L}\textbf{(}\theta+\epsilon^{*}\textbf{)}$.

Uncertainty Estimation. Although the high prediction performance of our annotator can be obtained, the labels may still be mispredicted due to the ambiguity of the face images (see an example in Fig. 3). Therefore, it is crucial to provide an uncertainty score for each prediction from the annotator. To this end, inspired by [46], we incorporate dropout techniques at each layer of MLP for uncertainty estimation in testing. This involves performing $k$ stochastic forward passes for a given test image $X$, each with a unique dropout pattern. So, we can obtain $k$ distinct softmax outputs for each demographic attribute $a$, denoted as $\{x^{1,(a)},...,x^{k,(a)}\}$. Then, the uncertainty score for $a$ on image $X$ (denoted as $V(X^{(a)})$) is calculated as $V(X^{(a)})=1-\Big{\{}\frac{1-\rho}{k}\sum_{i=1}^{k}x^{i,(a)}-\frac{\rho}{k^{2}}\sum_{i=1}^{k}\sum_{j=1}^{k}|x^{i,(a)}-x^{j,(a)}|\Big{\}}$, where $\rho\in[0,1]$ is a user-defined parameter to counterweight the measure of centrality (i.e., the first term in $\{\}$ indicates the likelihood of the prediction being correct) and dispersion (i.e., the second term in $\{\}$ reflects the consensus among the stochastic outputs).

Evaluation. To demonstrate our annotator’s effectiveness, we will answer the following questions: Q1: How are the general performance and generalization capability of our annotator compared with the baselines? Q2: How does sample difficulty affect the annotator’s performance? In leveraging the good generalization capabilities of CLIP, our annotator is trained on the VGGFace2 [47] dataset, which contains 9.1K individuals with 3.3M images. More importantly,  [48] provides comprehensive demographic annotations for this dataset. We compare our annotator with the current state-of-the-art face attribute prediction tools Face++ [42] and InsightFace [43]. Since they do not offer predictions for the race attribute, our evaluation is confined to gender and age. The mean and standard deviation are reported based on 5 random runs. More details are in Appendix A.1.1.

For Q1, Setting: We perform intra-domain (train on VGGFace2, test on its official test set) and cross-domain (train on VGGFace2, test on four AI-generated face datasets) evaluations. Specifically, A-FF++, A-DFDC, A-DFD, and A-Celeb-DF-v2 are selected from [19] for cross-domain evaluation. Since A-DFD and A-Celeb-DF-v2 have limited age and race annotations, our evaluation of these two is confined to gender. These datasets are chosen because they closely match our objective and are not used to train Face++ and InsightFace. Results: The ‘All’ results in Table 3 demonstrate our annotator’s superiority in general performance and generalization capability against Face++ and InsightFace. Under intra-domain evaluation, it surpasses the second-best method, Face++, by 5.8% on gender and 18.9% on age. In cross-domain evaluation, our annotator maintains high accuracy on all datasets, reflecting good generalization. Remarkably, on the A-FF++ dataset, our annotator outperforms Face++ by a substantial margin of up to 11.4% and InsightFace by 16.1% on age.

For Q2, Setting: We also design a stratified evaluation method by separating each test dataset into three subsets—Easy, Medium, and Hard based on the estimated uncertainty scores. Specifically, for each demographic attribute $a$, we define two thresholds $t_{1}^{a}$ and $t_{2}^{a}$, where $t_{1}^{a}<t_{2}^{a}$ (more details are in Appendix A.1.2). Then, we have $Easy={\{X\mid V(X^{(a)})<t_{1}^{a}\}}$, $Medium=\{X\mid t_{1}^{a}\leq V(X^{(a)})\leq t_{2}^{a}\}$, and $Hard=\{X\mid V(X^{(a)})>t_{2}^{a}\}$. Next, we sample 1,500 images from each subset. This stratification is crucial for a thorough examination of the model’s performance across a broad spectrum of data challenges. To avoid attribute-specific biases, each subset is balanced with respect to attribute. Results: Table 3 illustrates that while all methods show decreased accuracy as the sample difficulty level increases, our annotator demonstrates greater resilience. For example, under intra-domain evaluation, our annotator’s gender performance drops by 10.2% from easy to medium difficulty, compared to Face++’s 21.7% drop. In cross-domain scenario, our annotator experiences a 14.3% reduction on gender in A-Celeb-DF-v2 [19], versus InsightFace’s 16.2% from easy to hard.

Data Collection. We build our AI-Face dataset by collecting and integrating public AI-generated face images sourced from academic publications, GitHub repositories, and commercial tools. More details are in Appendix A.2.1. Specifically, the fake face images in our dataset originate from 4 Deepfake Video datasets (i.e., A-FF++ [19], A-DFDC [19], A-DFC [19], and A-Celeb-DF-v2 [19]), generated by 10 GAN models (i.e., AttGAN [49], MMDGAN [50], StarGAN [49], StyleGANs [49, 51, 52], MSGGAN [50], ProGAN [53], STGAN [50], and VQGAN [54]), and 8 DM models (i.e., DALLE2 [55], IF [55], Midjourney [55], DCFace [56], Latent Diffusiin [57], Palette [58], Stable Diffusion v1.5 [59], Stable Diffusion Inpainting [59]). This constitutes a total of 1,245,660 fake face images in our dataset. These fake images are correspondingly generated from 8 real source datasets (i.e., FFHQ [6], CASIA-WebFace [37], IMDB-WIKI [38], CelebA [36], and real images from FF++ [2], DFDC [39], DFD [40], and Celeb-DF-v2 [41]). This constitutes a total of 866,096 real face images in our dataset. In general, our dataset contains 30 subsets and 37 generation methods (i.e., 5 in A-FF++, 5 in A-DFD, 8 in A-DFDC, 1 in A-Celeb-DF-v2, 10 GANs, and 8 DMs). We use RetinaFace [60] for detecting and cropping faces to ensure each image only contains one face.

Annotator Prediction. For our collected images, annotation generation is iterative, integrating uncertainty scores into each prediction by our annotator in Phase 1, as shown in Fig. 2.

2) For the rest of the subsets, we identify the most ambiguous images that need human correction based on uncertainty scores. Specifically, for demographic attribute $a_{j}$ on subset $j$, we define a specific threshold $t^{a_{j}}$ (more details are in Appendix A.2.2). If $V(X^{(a_{j})})>t^{a_{j}}$, the annotation for attribute $a_{j}$ of the image $X$ will undergo a verification process, potentially requiring human re-annotation (see Fig. 3). In practice, we recruit three humans to correct the filtered images, consolidating their evaluations with a majority vote to finalize annotations.

Evaluation. To estimate our dataset’s quality, we will answer the following questions: Q1: Can we directly incorporate the existing annotations into our dataset? Q2: How is the effectiveness of human correction? Q3: How is the overall annotations’ quality of our dataset?

For Q1, Setting: We compare our dataset’s annotation quality before human correction on A-FF++ (i.e., Ours-FF++ (w/o Correction)) and A-DFDC (i.e., Ours-DFDC (w/o Correction)) against their existing annotation from [19]. We regard human re-labeled annotations as the ground truth. Results: The results in Table 4 ‘For Q1’ show superior annotation accuracy of our datasets. For example, Ours-FF++ (w/o Correction) surpasses A-FF++ by 83.97% in gender accuracy, and Ours-DFDC (w/o Correction) exceeds A-DFDC by 59.55%. The large performance indicates that identified images by annotation inconsistency are mislabeled in A-FF++ [19] and A-DFDC [19], and thus cannot be directly merged into our dataset. Some examples are shown in Appendix A.2.3.

For Q2, Setting: We consider two dataset versions: 1) Ours (w/o Correction), where annotations are not corrected by humans. 2) Ours, where annotations are corrected by humans. With the help of the uncertainty score, we sample 1,200 attribute-balanced images (400 easy, 400 medium, and 400 hard) from the whole dataset to ensure a fair evaluation. Three humans re-annotated these images to establish ground truth. Results: Table 4 ‘For Q2’ shows that human corrections improve performance across all attributes, increasing accuracy by 1.42% for gender, 0.92% for age, and 1.42% for race, validating the effectiveness of our correction strategy. More results see Appendix A.2.4.

For Q3, Setting: We randomly sample 1,200 images from the whole dataset. Three humans also re-annotated these images to create ground truth. Results: As shown in Table 4 ‘For Q3’, Ours reflects the approximate overall annotation quality of our dataset. Notably, the annotations of gender and race attributes show high correctness (e.g., 98.6667% ACC on gender and 86.2500% ACC on race). However, the age annotation shows a lower accuracy since it is challenging to differentiate.