Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,85 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
tags:
|
| 4 |
+
- deepfake-detection
|
| 5 |
+
- ai-image-detection
|
| 6 |
+
- dfd-arena
|
| 7 |
+
- bitmind
|
| 8 |
+
library_name: transformers
|
| 9 |
+
pipeline_tag: image-classification
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# PixelPrism v0.1 — DFD Arena Submission
|
| 13 |
+
|
| 14 |
+
Sanitized single-detector submission for the [BitMind Deepfake Detection Arena](https://huggingface.co/spaces/bitmind/dfd-arena-leaderboard).
|
| 15 |
+
|
| 16 |
+
This repo represents the **most informative single component** of PixelPrism's
|
| 17 |
+
production [V9 16-detector ensemble](https://pixelprism.ai/leaderboard), wrapped
|
| 18 |
+
in the BitMind `DeepfakeDetector` interface so it can be evaluated alongside
|
| 19 |
+
NPR / UCF / CAMO on the public leaderboard.
|
| 20 |
+
|
| 21 |
+
## What's in this submission
|
| 22 |
+
|
| 23 |
+
A wrapper around the **Swin V2 transformer head**
|
| 24 |
+
([haywoodsloan/ai-image-detector-deploy](https://huggingface.co/haywoodsloan/ai-image-detector-deploy),
|
| 25 |
+
MIT licensed). In PixelPrism's V9 permutation-importance audit (8000 samples,
|
| 26 |
+
5 reps), Swin V2 ranked **#1 by a wide margin** at importance **0.271**, vs
|
| 27 |
+
0.109 for the next-best detector (vit3) and 0.032 for DIRE-FLUX. It alone
|
| 28 |
+
accounts for ~38% of V9's total discriminative power.
|
| 29 |
+
|
| 30 |
+
## What's NOT in this submission
|
| 31 |
+
|
| 32 |
+
The full PixelPrism V9 ensemble fuses **16 detectors** via a
|
| 33 |
+
`HistGradientBoostingClassifier` meta-classifier:
|
| 34 |
+
|
| 35 |
+
```
|
| 36 |
+
fft, vit, vit2, vit3, dire (SD 1.5), clip, srm, exif, face,
|
| 37 |
+
cfa, prnu, c2pa, anatomy, swin, dire_sdxl, dire_flux
|
| 38 |
+
```
|
| 39 |
+
|
| 40 |
+
Some V9 components depend on weights that are not MIT-redistributable:
|
| 41 |
+
- `dire_flux` uses FLUX.1-schnell (non-commercial license)
|
| 42 |
+
- `dire_sdxl` uses Stability SDXL (CreativeML OpenRAIL-M)
|
| 43 |
+
- `face` uses FaceForensics++ Xception variants (access-gated)
|
| 44 |
+
|
| 45 |
+
Those stay in our internal production stack rather than the public submission.
|
| 46 |
+
|
| 47 |
+
## Live full-ensemble numbers
|
| 48 |
+
|
| 49 |
+
The full V9 ensemble is live at **<https://pixelprism.ai/api/detect>** (paid)
|
| 50 |
+
and **<https://pixelprism.ai/api/scan-public>** (5/day free tier). Per-generator
|
| 51 |
+
detection rates and 30-day drift trend are published at
|
| 52 |
+
**<https://pixelprism.ai/leaderboard>** (refreshed monthly with each retrain).
|
| 53 |
+
|
| 54 |
+
V9 internal holdout (8000 stratified samples, 4000 real / 4000 AI):
|
| 55 |
+
|
| 56 |
+
| Metric | V9 |
|
| 57 |
+
|---|---|
|
| 58 |
+
| Overall | 96.7% |
|
| 59 |
+
| Real | 96.1% |
|
| 60 |
+
| AI | 97.4% |
|
| 61 |
+
| Per-generator min | 91.0% (Grok) |
|
| 62 |
+
| Drift gap (fresh AI vs known AI) | −2.7pp (fresh AI now BEATS known AI) |
|
| 63 |
+
|
| 64 |
+
## Files in this repo
|
| 65 |
+
|
| 66 |
+
| File | Purpose |
|
| 67 |
+
|---|---|
|
| 68 |
+
| `pixelprism_detector.py` | The `DeepfakeDetector` subclass registered as `PixelPrism` in `DETECTOR_REGISTRY` |
|
| 69 |
+
| `pixelprism_config.yaml` | YAML config with `hf_repo`, `backbone_repo`, `ai_label_idx` |
|
| 70 |
+
| `model.safetensors` | Swin V2 weights (re-hosted, byte-identical to upstream) |
|
| 71 |
+
| `config.json` | Swin V2 model config |
|
| 72 |
+
| `preprocessor_config.json` | Swin V2 image preprocessor config |
|
| 73 |
+
| `README.md` | This file |
|
| 74 |
+
|
| 75 |
+
## Citation / contact
|
| 76 |
+
|
| 77 |
+
If you use this in research or a comparison study, cite:
|
| 78 |
+
- The Swin V2 detector: [haywoodsloan/ai-image-detector-deploy](https://huggingface.co/haywoodsloan/ai-image-detector-deploy)
|
| 79 |
+
- PixelPrism's full ensemble methodology: <https://pixelprism.ai/leaderboard>
|
| 80 |
+
|
| 81 |
+
Operator: Chris Crawley, PixelPrism.ai · <https://pixelprism.ai>
|
| 82 |
+
|
| 83 |
+
## License
|
| 84 |
+
|
| 85 |
+
MIT (matches the upstream Swin V2 model + matches the BitMind DFD Arena requirement).
|