The increasing prevalence of AI images on social media creates new problems in detecting these manipulated images. Modern techniques to detect AI-generated images are often implemented after images’ distribution, which makes it hard to control the distribution of such content. Also, there is the possibility that the performance of modern detection algorithms varies depending on demographics. In this paper, we propose a fairness-aware framework for detecting AI images before posting them on social media. For implementing the method, we chose EfficientNet-B4 and ResNet-50 models as the detection backbones and used the DAID fairness framework for fine-tuning. The threshold was chosen using Youden’s J statistic to address the domain shift problem. We conducted experiments on 95,634 cross-compressed DFDC images across six demographics. The ensemble provided an ROC-AUC of 0.6395 with demographic skew of 0.0184 and subgroup AUCs of 0.6298–0.6482.

AI-generated images, deepfake detection, preupload detection, demographic fairness, DAID, domain shift, social media security