Identifying Diffusion Model Generated Synthetic Faces Covering OSN Context and Ethnic Diversity

Today's advancements in technology have democratized the creation of highly realistic visual simulations, posing significant challenges to the integrity of digital content. Synthetic faces generated through sophisticated AI models such as generative adversarial networks and diffusion models can convincingly deceive even the most experienced observers. These digital fabrications are being used increasingly to spread false information on social media and other online platforms. Although researchers have started investigating unique patterns and artifacts introduced by such generator models, so as to detect synthetic images, the efficacy of these detection techniques is compromised by the repetitive sharing and processing of images on social media platforms. Operations such as compression applied during these processes gradually degrade the forensic traces essential for accurate detection. This work deals with the problem of detecting synthetic faces generated by diffusion models. To address the “synthetic-vs-real” problem in practical settings, we create a dataset of six different SOA diffusion model generated realistic facial images, along with an OSN compressed version of the dataset, wherein we include four prevalent online social platforms from present times. We propose a solution by utilizing the attention mechanism, which allows us to capture dependencies more efficiently, often requiring fewer parameters to achieve similar or better performance compared to networks without attention. Even after working with a reduced set of parameters, our model achieves a high detection accuracy (average) ≥ 92%). Additionally, it achieves strong detection results on post-processed images, reflecting real-world OSN environments.