Comprehensive Deepfake Detection Dataset: Real and Synthetic Frames from Roop and Akool AI Technologies

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.

本数据集是面向深度伪造（deepfake）检测的前沿研究资源，共从480段视频中提取110694帧图像。其包含两大核心类别：深度伪造帧与真实帧。其中106948个深度伪造帧由450段视频生成，所用工具包括Roop换脸器（Roop Faceswapper）与Akool AI等先进人工智能工具；而3746个真实帧则源自30段原始真实视频。本数据集经过严格遴选与整理，确保了样本的多样性、平衡性与高质量代表性，是训练与评估深度伪造检测模型的宝贵资源。

本数据集的采集流程已获得水仙国际大学（Daffodil International University）的伦理审查批准，确保符合深度伪造检测系统领域数据采集与研究的伦理规范。

核心特性：
总视频数：480段（其中深度伪造视频450段，真实视频30段）。
帧分布：深度伪造帧106948个，真实帧3746个。
深度伪造生成工具：采用Roop换脸器（Roop Face-Swapper）与Akool AI等前沿技术生成合成视频。
人口统计学多样性：涵盖15名男性与15名女性的面部帧，确保在面部特征、光照条件与拍摄环境上具备丰富的多样性。
数据集平衡性：经过精心整理，确保模型训练与评估的公平性。
应用场景：适用于开发网络安全、数字取证与媒体真实性核验领域的可扩展实时检测系统。

本数据集为探索与优化深度伪造检测模型提供了绝佳契机，可凭借其多样化且高质量的基准数据集，应对合成媒体带来的各类挑战。