Deepfake Database

Data Collection and Processing

• Face Extraction and Alignment: All faces were extracted using the Viola-Jones detector and aligned using a trained neural network for facial landmark detection. To ensure a balanced distribution of faces, the number of selected frames for extraction per video is proportional to the number of camera angles and illumination changes on the target face. On average, approximately 50 faces were extracted per scene.
• Dataset Composition: The dataset includes both forged and real face images. Real faces were also extracted from various internet sources, maintaining the same resolution as the forged faces. The dataset was manually reviewed to remove misaligned and incorrectly detected faces. Both classes (real and forged faces) were balanced for good-resolution and poor-resolution images to avoid classification bias.

Key Features

• Diversity: The dataset includes a wide variety of faces from different videos to ensure diversity.
• Real-world Conditions: Videos are compressed using different levels of the H.264 codec, replicating real-world conditions.
• Balanced Distribution: The number of faces extracted per video is balanced based on camera angles and illumination changes, ensuring a representative sample.
• Manual Review: The dataset was manually reviewed to remove any misalignments and incorrect face detections, ensuring high-quality data.
• Validation Set: Includes a validation set for model testing and evaluation.

Enhanced Content and Applications

This dataset can be used for:

• Deepfake Detection: Training and evaluating models designed to detect forged videos.
• Face Recognition: Enhancing the robustness of face recognition systems against deepfake attacks.
• Video Forensics: Assisting in the development of forensic tools to identify and analyze video manipulations.
• AI Ethics Research: Studying the ethical implications and potential misuse of deepfake technology.