Vi0509/kaeva-deepfake-datasets

Kaeva Deepfake Detection — Training Datasets (V1–V9)

This repository documents all training datasets used across Kaeva deepfake detection model versions V1 through V9. No raw data is hosted here — this serves as a comprehensive reference card.

Training code: Viraj-FG/kaeva-verify/training/

---

Dataset Inventory

Established Benchmarks

Dataset	Type	Source	License
CIFAKE	Real + AI-generated (CIFAR-10 scale)	HF: Bird/CIFAKE	CC BY-SA 4.0
ArtiFact	Multi-generator forensics benchmark	GitHub: awsaf49/artifact	Research
OpenFake	Open-source deepfake benchmark	GitHub	Research
DeepFakeFace	Face-swap deepfakes	Kaggle	Research
GenImage	Multi-generator image detection	GitHub: GenImage-Dataset	Research
Kaggle DFD	Deepfake Detection Challenge	Kaggle DFD	Competition

Face Datasets (Real Baselines)

Dataset	Description	Source	License
CelebA-HQ	30k high-quality celebrity faces	GitHub: tkarras/progressive_growing_of_gans	Non-commercial research
FFHQ	70k Flickr-sourced high-quality faces	GitHub: NVlabs/ffhq-dataset	CC BY-NC-SA 4.0

Large-Scale Image Datasets

Dataset	Description	Source	License
ImageNet-1k	1.28M images, 1000 classes	image-net.org	Research (non-commercial)
ai-artbench	AI-generated art benchmark	HF: ramonpzg/ai-artbench	MIT
dima806/ai_vs_real	AI vs real photo classification	HF: dima806/ai_vs_real	CC BY 4.0

Web-Scraped Sources

Source	Type	Usage
thispersondoesnotexist.com	GAN-generated faces (StyleGAN)	Fake samples
picsum.photos	Random real photographs	Real baseline samples
StyleGAN3	NVIDIA StyleGAN3 generated faces	Fake samples (GAN family)

---

V9 Generator Coverage

V9 expanded coverage to 10 modern generators to ensure broad generalization:

Generator	Family	Notes
sdxl_turbo	Stable Diffusion XL Turbo	Distilled, few-step
playground_v2.5	Playground AI	Aesthetic-optimized
pixart_sigma	PixArt-Σ	DiT-based
kandinsky3	Kandinsky 3	Sber AI
sd35_medium	Stable Diffusion 3.5 Medium	MMDiT
kolors	Kolors (Kwai)	Chinese text-to-image
sd35_large	Stable Diffusion 3.5 Large	MMDiT (large)
flux_schnell	FLUX.1 [schnell]	Black Forest Labs, distilled
flux_dev	FLUX.1 [dev]	Black Forest Labs, guidance-distilled
wan2.1	Wan 2.1	Video/image generation

---

Data Principles

1. Real Baseline — Pristine

All real images are sourced at highest available quality with no re-compression. This ensures the model learns authentic camera/sensor characteristics rather than compression artifacts.

2. Compression Washing for Fakes

Fake images undergo compression washing (JPEG re-save at varying quality levels, WebP conversion, etc.) to strip superficial generation artifacts. This forces the model to detect deeper structural signals rather than relying on compression-level shortcuts.

3. GER Buffer — Hard Negatives

A Generator-Error-Rate (GER) buffer of hard negative samples is maintained. These are AI-generated images that closely mimic real image statistics and are difficult to classify. Including them during training improves calibration and pushes the decision boundary into the ambiguous region where it matters most.

---

Training Scripts

All training code is maintained in the private repository:

Viraj-FG/kaeva-verify/training/
├── train_lnclip.py          # LNCLIP LayerNorm probe training
├── train_audio.py           # Audio deepfake detector training
├── data_pipeline.py         # Dataset loading & augmentation
├── compression_wash.py      # Compression washing transforms
└── ger_buffer.py            # GER hard negative mining

---

Citation

If you use this dataset documentation or the Kaeva models, please reference:

@misc{kaeva2026,
  title={Kaeva: Multi-Modal Deepfake Detection},
  author={Viraj},
  year={2026},
  url={https://github.com/Viraj-FG/kaeva-verify}
}