siddharthksah/DeepSafe-benchmark
收藏资源简介:
# DeepSafe Multi-Modal Deepfake Detection Benchmark A balanced, multi-modal benchmark dataset for evaluating deepfake detection systems across images, audio, and video. ## Dataset Summary | Modality | Real | Fake | Total | |----------|------|------|-------| | Images | 2,000 | 2,000 | 4,000 | | Audio | 1,000 | 1,000 | 2,000 | | Video | 100 | 100 | 200 | | **Total** | **3,100** | **3,100** | **6,200** | ## Structure ``` images/ real/{source}/{id}.{ext} fake/{generator}/{id}.{ext} audio/ real/{source}/{id}.wav fake/{generator}/{id}.wav video/ real/{source}/{id}.mp4 fake/{generator}/{id}.mp4 metadata.json ``` ## Generators Covered ### Images (34+ generators) **Real sources:** COCO, OpenFake **Fake generators:** DALL-E 2, DALL-E 3, Midjourney v5/6/7, Stable Diffusion 1.3/1.4/2/3/XL, Flux 1, GPT Image 1, Grok 2, Imagen 3/4, Ideogram 2/3, Firefly, GLIDE, Recraft v2/v3, HiDream, Aurora, Chroma, Lumina, Mystic, Halfmoon, Frames, and more. ### Audio (25+ generators) **Real sources:** ASVspoof, LibriSpeech, LJSpeech, VCTK, In-the-Wild, CodecFake **Fake generators:** HiFiGAN, MelGAN, WaveGlow, Tacotron, Griffin-Lim, Parallel WaveGAN, Full-Band MelGAN, Conformer FastSpeech2, Neural Codec, WaveNet, ASVspoof attack types (A01-A06), and more. ### Video (15+ generators) **Real sources:** MSRVTT, LAVDF **Fake generators:** Sora, Gen-2, Moonvalley, MorphStudio, CRaFT-er, LaVie, ModelScope, Show-1, Hotshot, WildScrape, LAVDF manipulations (audio/video/both), and more. ## Metadata Format `metadata.json` contains one entry per file: ```json { "id": "img_00001", "path": "images/real/coco/img_00001.jpg", "modality": "images", "label": "real", "generator": "coco", "format": "jpg" } ``` ## Sampling Methodology Files are stratified-sampled across generators using round-robin allocation to ensure balanced representation. Each generator contributes proportionally to its available pool size, preventing over-representation of any single source. ## License This dataset aggregates samples from publicly available research datasets. Each source dataset retains its original license terms. This compilation is provided for research and benchmarking purposes. ## Citation If you use this dataset, please cite: ``` @misc{deepsafe_benchmark_2026, title={DeepSafe Multi-Modal Deepfake Detection Benchmark}, year={2026}, url={https://deepsafe.fyi} } ```
# DeepSafe多模态深度伪造(deepfake)检测基准数据集 本数据集为平衡式多模态基准数据集,用于在图像、音频与视频维度上评估深度伪造检测系统。 ## 数据集概览 | 模态 | 真实样本数 | 伪造样本数 | 总计 | |----------|------|------|-------| | 图像 | 2,000 | 2,000 | 4,000 | | 音频 | 1,000 | 1,000 | 2,000 | | 视频 | 100 | 100 | 200 | | **总计** | **3,100** | **3,100** | **6,200** | ## 数据集结构 images/ real/{来源}/{编号}.{文件扩展名} fake/{生成模型}/{编号}.{文件扩展名} audio/ real/{来源}/{编号}.wav fake/{生成模型}/{编号}.wav video/ real/{来源}/{编号}.mp4 fake/{生成模型}/{编号}.mp4 metadata.json ## 覆盖的生成模型 ### 图像模态(覆盖34+种生成模型) **真实样本来源:** COCO、OpenFake **伪造生成模型:** DALL-E 2、DALL-E 3、Midjourney v5/6/7、Stable Diffusion 1.3/1.4/2/3/XL、Flux 1、GPT Image 1、Grok 2、Imagen 3/4、Ideogram 2/3、Firefly、GLIDE、Recraft v2/v3、HiDream、Aurora、Chroma、Lumina、Mystic、Halfmoon、Frames 以及更多。 ### 音频模态(覆盖25+种生成模型) **真实样本来源:** ASVspoof、LibriSpeech、LJSpeech、VCTK、In-the-Wild、CodecFake **伪造生成模型:** HiFiGAN、MelGAN、WaveGlow、Tacotron、Griffin-Lim、Parallel WaveGAN、Full-Band MelGAN、Conformer FastSpeech2、Neural Codec、WaveNet、ASVspoof攻击类型(A01-A06)以及更多。 ### 视频模态(覆盖15+种生成模型) **真实样本来源:** MSRVTT、LAVDF **伪造生成模型:** Sora、Gen-2、Moonvalley、MorphStudio、CRaFT-er、LaVie、ModelScope、Show-1、Hotshot、WildScrape、LAVDF操纵(音频/视频/二者同时)以及更多。 ## 元数据格式 `metadata.json` 为每个样本文件对应一条条目: json { "id": "img_00001", "path": "images/real/coco/img_00001.jpg", "modality": "images", "label": "real", "generator": "coco", "format": "jpg" } ## 采样方法 本数据集采用分层抽样策略,通过轮询分配方式对各生成模型的样本进行采样,以确保样本分布均衡。各生成模型的样本占比与其可用样本池规模成正比,避免单一来源样本占比过高。 ## 许可协议 本数据集汇总自公开的学术研究数据集,各原始数据集仍保留其原有许可协议条款。本数据集仅用于学术研究与基准测试用途。 ## 引用方式 若您使用本数据集,请引用如下文献: @misc{deepsafe_benchmark_2026, title={DeepSafe Multi-Modal Deepfake Detection Benchmark}, year={2026}, url={https://deepsafe.fyi} }




