README.md

metadata

license: other
license_name: mixed-upstream
extra_gated_prompt: >-
  This dataset aggregates media from ~20 upstream sources with differing terms.
  Several (ASVspoof, DF40, LAV-DF, MSR-VTT, ImageNet, MLAAD) are research-only,
  EULA-bound, or not redistributable. Accepting here does NOT relieve you of
  complying with each upstream source's own licence.

  The audio and video splits contain recordings of real people. Treat them as
  biometric data and handle accordingly.
extra_gated_fields:
  I will use this dataset for non-commercial research only: checkbox
  I will comply with each upstream source's own licence: checkbox
  I understand this contains biometric data: checkbox
task_categories:
  - audio-classification
  - image-classification
  - video-classification
tags:
  - deepfake-detection
  - multimodal
  - benchmark
size_categories:
  - 10K<n<100K
pretty_name: DeepSafe Evaluation Dataset

DeepSafe Evaluation Dataset

Evaluation set for DeepSafe, a deepfake detection benchmark.

Tiers

Tier	Samples	Generators	Size	Use
master_eval_small/	198	116	1.7 GB	smoke test, under 2 min
master_eval/	15,454	411	10 GB	the standard benchmark
master_eval_full/	45,954	411	25 GB	complete set

Medium tier composition: 9,954 image, 3,500 audio, 2,000 video.

from huggingface_hub import snapshot_download
snapshot_download("deepsafe/evaluation-dataset", repo_type="dataset",
                  allow_patterns="master_eval_small/**", local_dir="./data")

What the data shows

Measured on this set, an ensemble of 19 published detectors catches 66.2% of fakes at a 6.5% false positive rate, and the failures are concentrated:

Generator	Caught
Hunyuan (video)	1.7%
Sora (video)	7.0%
Veo (video)	8.0%
Stable Diffusion 1.4 (image)	100%
Midjourney v5 (image)	100%

Detectors are strong on the 2022-2023 diffusion models the literature was built around and weak on recent video generators. That contrast is the reason this set exists.

Upstream sources

Freely redistributable: LibriSpeech (CC-BY-4.0), LJSpeech (public domain), VCTK (CC-BY-4.0), Common Voice (CC-0), COCO (CC-BY-4.0).

Research-only, EULA-bound, or not redistributable: ImageNet, ASVspoof, DF40, LAV-DF, MSR-VTT, MLAAD.

This repository is gated because of the second group. Each upstream source remains governed by its own terms, which you must satisfy independently.

Rebuilding from source

The manifests (metadata.json per tier) record every sample's origin, so the set can be reconstructed from the original sources rather than from this mirror. See eval/build_master_eval.py in the DeepSafe repository.

Takedown

If you hold rights to any included material and want it removed, open an issue at https://github.com/deepsafehq/deepsafe-bench. Removal within 48 hours, no questions asked.

Caveat

The manifests carry no language field. Do not infer multilingual performance from this set without adding language labels first.