DeepFakeFace

The rise of deepfake images, especially of well-known personalities, poses a serious threat to the dissemination of authentic information. To tackle this, we present a thorough investigation into how deepfakes are produced and how they can be identified. The cornerstone of our research is a rich collection of artificial celebrity faces, titled DeepFakeFace (DFF). We crafted the DFF dataset using advanced diffusion models and have shared it with the community through online platforms. This data serves as a robust foundation to train and test algorithms designed to spot deepfakes. We carried out a thorough review of the DFF dataset and suggest two evaluation methods to gauge the strength and adaptability of deepfake recognition tools. The first method tests whether an algorithm trained on one type of fake images can recognize those produced by other methods. The second evaluates the algorithm's performance with imperfect images, like those that are blurry, of low quality, or compressed. Given varied results across deepfake methods and image changes, our findings stress the need for better deepfake detectors. Our DFF dataset and tests aim to boost the development of more effective tools against deepfakes.

We present a new dataset named DeepFakeFace (DFF) to assess the ability of deepfake detectors to distinguish AI-generated and authentic images. There are 30,000 pairs of real and fake images. Since we aim to protect the privacy of celebrities, 30,000 real images of dataset all comes from IMDB-WIKI dataset. The dataset consists of 120,000 images which incorporate 30,000 real images and 90,000 fake images.  We employ three different generative models for synthesizing deepfakes: Stable Diffusion v1.5, Stable Diffusion Inpainting and a powerful toolbox InsightFace. Each model generates 30,000 fake images.

Diffusion models, which create high-resolution images via the sequential deployment of denoising autoencoders, are an integral part of our methodology. Direct pixel-level operation, however, proves resource-intensive in terms of time and computational complexity. To counteract this, stable diffusion harnesses diffusion models within the latent space. This not only conserves computational resources but also maintains the quality and flexibility of generated images. With its prowess in synthesizing photo-realistic images from any given input text, we adopted stable diffusion for crafting deepfakes. These images bear a resolution of 512 × 512. Our study utilizes both the Stable Diffusion v1.5 and Stable Diffusion Inpainting models. Additionally, for a multifaceted approach, the InsightFace toolbox—equipped with top-tier algorithms for face recognition, detection, alignment, and swapping—also contributes to our deepfake generation.

The IMDB-WIKI dataset, known for its extensive compilation of face images annotated with gender and age, is our primary source of authentic images. Leveraging this dataset allows for effortless extraction of gender and age metadata from its label files. For consistency, images are randomly matched based on gender, and this configuration is adhered to in subsequent methodologies.  Upon retrieval of gender, age, and identity for each image, prompts corresponding to each image are generated. These prompts adhere to the template: "name, celebrity, age", where "name" and "age" are replaced by the image's actual identity and age, respectively. Though the IMDB-WIKI dataset furnishes aligned faces with the original facial bounding box, discrepancies in accuracy were noted in some bounding boxes. To address this, we utilized the cutting-edge RetinaFace face detector to redefine facial bounding boxes and generate corresponding mask images.  Equipped with this refined data, deepfakes are then generated using Stable Diffusion v1.5, Stable Diffusion Inpainting, and InsightFace, respectively.

4. Experiments