Thiru1122's picture
Add UnivFD-only benchmark framework
7591a5b
|
Raw History Blame Contribute Delete
1.64 kB
# Benchmark Dataset Directory
This directory provides an isolated, standardized benchmark harness for testing deepfake detectors.
## Directory Structure
```text
benchmark/
β”œβ”€β”€ README.md
β”œβ”€β”€ metadata.json
β”œβ”€β”€ images/
β”‚ β”œβ”€β”€ real/ # Ground-truth photographic images
β”‚ β”œβ”€β”€ ai_generated/ # Ground-truth synthetically generated images
β”‚ β”œβ”€β”€ manipulated/ # Ground-truth edited/spliced images
β”‚ └── corrupted/ # Corrupted or malformed files for rejection testing
└── results/ # Benchmark execution outputs (JSON and CSV)
```
## Input Availability & Addition Protocol
In accordance with strict repository isolation rules:
- No files were copied from local workspaces or external paths outside this Space repository.
- No ground-truth labels were fabricated.
### How to Manually Add Benchmark Images
1. Place natural photographs into `benchmark/images/real/`.
2. Place AI-generated images (e.g., from Diffusion, GAN, or Flow-Matching models) into `benchmark/images/ai_generated/`.
3. Place edited or composite images into `benchmark/images/manipulated/`.
4. Place truncated or corrupted test files into `benchmark/images/corrupted/`.
5. Update `benchmark/metadata.json` with the filename, category, and verified ground-truth origin. If ground truth cannot be verified with documentary proof, mark `ground_truth: "UNKNOWN"`.
## Non-Authenticity Disclaimer
This benchmark framework is an operational diagnostic tool. Successful execution does not imply detector accuracy, and consensus across models does not verify photographic authenticity.