Comprehensive Deepfake Detection Dataset: Real and Synthetic Frames from Roop and Akool AI Technologies
Description
This dataset is a cutting-edge resource for deepfake detection, containing 110,694 frames extracted from 480 videos. It features two primary categories: deepfake and real frames. The deepfake frames (106,948) were generated from 450 videos using advanced AI tools such as Roop Faceswapper and Akool AI, while the real frames (3,746) were derived from 30 authentic videos. The dataset is meticulously curated to ensure diversity, balance, and high-quality representation, making it an invaluable resource for training and evaluating deepfake detection models. The dataset collection process was conducted with ethical approval from Daffodil International University, ensuring adherence to ethical standards for data collection and research in deepfake detection systems. Key Features: Total Videos: 480 (450 deepfake, 30 real). Frame Distribution: 106,948 deepfake frames and 3,746 real frames. Deepfake Generation Tools: Leveraged state-of-the-art technologies like Roop Face-Swapper and Akool AI for synthetic video creation. Demographic Diversity: Includes frames from 15 males and 15 females, ensuring varied representation across facial features, lighting conditions, and environments. Balanced Dataset: Carefully curated for fair model training and evaluation. Applications: Ideal for developing scalable, real-time detection systems in cybersecurity, digital forensics, and media integrity verification. This dataset offers an unparalleled opportunity to explore and enhance deepfake detection models, addressing the challenges posed by synthetic media with a diverse and high-quality benchmark.
Files
Steps to reproduce
Steps to reproduce Data Collection Process Outline: 1. Overview of Data Collection: The dataset was created to address the growing challenge of deepfake detection by providing a balanced and diverse set of real and deepfake frames. The data comprises 480 videos: 450 used to generate deepfake frames and 30 real videos for authentic frames. The entire process was designed to ensure reproducibility and transparency. 2. Methods and Protocols: a. Video Selection and Preparation Subjects: 30 individuals (15 males and 15 females) were chosen based on diverse demographics. Source Videos: Obtained from controlled recordings to ensure clarity and consistency in lighting and backgrounds. Additional video content, such as podcasts and interviews, was sourced from platforms like YouTube and other publicly available media repositories. b. Deepfake Video Generation Tools Used: Roop Face-Swapper: A deep learning-based tool used to swap faces in videos with high precision. Akool AI: A generative AI tool that creates high-quality deepfake videos for commercial and research applications. Process: For each individual, 15 real videos were used to generate deepfake videos for all other individuals in the dataset, resulting in 450 deepfake videos. Parameters like frame rate, resolution, and face-swapping precision were optimized for quality. 3. Frame Extraction Software Used: OpenCV (Python library) for video processing. Method: Frames were extracted at a consistent interval of 5 frames per second (FPS). Real and deepfake frames were stored in separate directories, categorized by gender. 4. Face Detection and Cropping Algorithm: Multi-Task Cascaded Convolutional Networks (MTCNN). Steps: Detected faces in each frame and cropped them to a standard size (500x500 pixels). Ensured only high-quality frames with clearly detected faces were included. 5. Data Preprocessing Normalization:Pixel values of images were normalized to the range [0, 1]. Labels assigned: 0 for real frames and 1 for deepfake frames. Augmentation: Random rotations, flips, and scaling were applied using torchvision.transforms to enhance model robustness. 6. Software and Workflows Programming Language: Python Key Libraries and Tools: OpenCV: For video processing and frame extraction. MTCNN: For face detection and cropping. PyTorch: For preprocessing and augmentation. Roop and Akool AI: For generating deepfake videos. 7. Reproducibility: Codebase: Scripts for video processing, frame extraction, and preprocessing were developed in Python and can be shared for replication. Hardware Requirements: GPU-enabled systems were used for deepfake generation and face detection to ensure efficiency.