34data/sowaiba01-deepfake-fake

sowaiba01-deepfake-fake

Mirror of the exact bmcore v24 holdout subset. Source: Sowaiba01/Deepfake.

All 5,426 local fake basenames match the public fake/ subset. Face swaps remain semisynthetic.

Contains 5426 semisynthetic image files. This mirror repackages the media; it does not grant additional rights.

Source revision reviewed: d8b189507fc026b3e0c794a06c318cb13d0c8743.

Original source card

license: mit task_categories:

• —image-classification language:
• —en tags:
• —deepfake
• —face-swap
• —detection
• —computer-vision
• —insightface size_categories:
• —10K<n<100K ---

DeepGuard Deepfake Dataset

A paired deepfake dataset for training and benchmarking deepfake detection models. Generated as part of the DeepGuard AI Project (2026).

📥 2,301+ all-time downloads

Dataset Statistics

• —Fake images: 5426 (face-swapped using InsightFace inswapper_128)
• —Real images: 5426 (original paired faces)
• —Total: 10852 images
• —Format: JPEG, high quality (95%)
• —Generation method: InsightFace inswapper_128 (ONNX runtime)

Why This Dataset is Unique

• —Paired structure: Every fake has a matching real image of the same target person
• —Modern method: inswapper_128 powers most 2024-2026 real-world deepfake apps
• —Traceable metadata: Source identity, target face, and method per image
• —Open access: No sign-up required, instant load_dataset() access

Quick Start

from datasets import load_dataset
ds = load_dataset('Sowaiba01/Deepfake')

Detection Model

Trained EfficientNet-B4 detection model also available:

from huggingface_hub import hf_hub_download
model_path = hf_hub_download('Sowaiba01/deepguard-ai', 'efficientnet_b4_deepguard.pth')

Validation accuracy: 91.54% on 140K Real & Fake Faces dataset.

Citation

@dataset{deepguard2026,
  author = {Arshad, Sowaiba},
  title  = {DeepGuard Deepfake Dataset},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Sowaiba01/Deepfake}
}