SegDeep Dataset: A Large-Scale Dataset and Benchmark for Semantic Segmentation of Diffusion-Based Facial Manipulations

The dataset contains 28,245 facial images organized into three classes: Authentic images, DeepFake manipulations, and Face Swap manipulations. Authentic samples were obtained from the FFHQ dataset, while manipulated samples were generated using Stable Diffusion-based models. Pixel-level annotations were automatically generated during the manipulation process using a face-parsing model to identify the affected facial regions. The utility of the dataset was evaluated using different semantic segmentation architectures.

This dataset is a derivative work based on the Flickr-Faces-HQ (FFHQ) dataset. Appropriate credit is provided to the original FFHQ authors, and modifications introduced during the construction of the proposed dataset are documented