README.md

metadata

title: DeepFakeScanner
sdk: docker
app_port: 7860
pinned: false
license: apache-2.0
short_description: AI-generated & manipulated image detection (v0.4.0)

DeepFakeScanner

A commercial deepfake / AI-generated image detection service.

What it is

A FastAPI web service that scans an uploaded image and returns a structured verdict: authentic, ai_generated, deepfake, edited, or uncertain — with per-class probabilities, per-detector signals, and a C2PA provenance check.

Live

	URL
Frontend	https://veridicate.com
API	https://api.veridicate.com
Source (HF mirror)	https://huggingface.co/spaces/veridicate/scanner

Status

Stage 2 deployed (v0.3.0-stage2). The CLIP classifier head was fine-tuned on a 100k commercially-licensed dataset (50k Open Images + 50k Flux.1-schnell) and is currently serving real predictions in production.

In-distribution test-set metrics (10k held-out images):

• Accuracy: 98.44%
• Precision (AI class): 98.25%, recall: 98.64%, F1: 98.44%

Stage 3A wired + verified 2026-05-30 (v0.4.0-stage3a); ships to production on merge of feat/akila-20260515 → main. Multi-generator dataset built: 50k Flux + 20k SDXL + 20k SD 3.5 Medium + 10k AuraFlow on the AI side, matched authentic from Open Images V7. SDXL held out of training so the heldout split is a true generalisation test. The trained head (head_v3a.pt) is published to the private HF Hub repo, the runtime config now defaults to it (config.py), and a filtered, test-gated GitHub Action deploys the inference service on merge. Headline verified numbers vs the Stage 2 baseline:

Split	Baseline	Candidate	Δ
heldout SDXL (UNSEEN in training)	83.93%	89.48%	+5.55 pp
test (in-distribution)	96.47%	98.43%	+1.96 pp
test_augmented (robustness)	93.96%	98.41%	+4.46 pp

The +5.55 pp on the SDXL holdout is the load-bearing number — SDXL was held entirely out of training, so it's the closest available proxy for how the model will behave on generators it never saw. The family-fingerprint approach (train on a diverse mix of open generators, inherit coverage of closed generators) is validated. Detailed audit trail in docs/stage3a-implementation.md.

Known limitation (until the Stage 3A merge deploys): the live model is still the Stage 2 head, trained on Flux.1-schnell only. Other generators (Gemini/Imagen 3, DALL-E 3, Midjourney, Grok, Stable Diffusion) are out-of-distribution for the currently-live model and it often returns uncertain verdicts. The Stage 3A head (queued to ship) closes most of this gap. Rollback is a one-line env override (MODEL_VERSION + HEAD_CHECKPOINT_HF_FILENAME) — both heads live in the same private HF Hub repo.

Roadmap at a glance

Stage	What it delivers	Status
1	Working website, API, deploy pipeline. Detector returns random guesses.	✅ Done
2	A trained classifier — 98% accurate on Flux-family AI images.	✅ Done, live (v0.3.0-stage2)
3A	Broad coverage across the AI image-generation landscape — CLIP head retrained on Flux + SDXL + SD 3.5 + AuraFlow.	✅ Wired + verified 2026-05-30 (v0.4.0-stage3a); ships on merge to main
3B	Frequency-artifact detector (FFT/DCT, generator-agnostic) brought online.	🔭 Queued after Stage 3A ships
4	Production scale: faster hosting, paid tier, user accounts.	⏸️ After Stage 3
5	Enterprise capability: licensed paid-API training data, face-swap detection, adversarial robustness.	🔭 Future

Full roadmap with per-stage strengths, weaknesses, and how each weakness gets fixed: docs/plan.md. Plain-English summary up front; technical detail below; glossary at the end for non-technical readers.

Quick start

# CPU PyTorch first (lean install)
pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu
pip install -r requirements.txt
pip install -e .

# Pre-download CLIP weights
python scripts/download_weights.py

# Run the server
uvicorn deepfake_scanner.api.v1:app --reload --port 7860

Then:

curl -F "file=@some_image.jpg" http://localhost:7860/v1/scan/image | jq

Or against the live API:

curl -F "file=@some_image.jpg" https://api.veridicate.com/v1/scan/image | jq

Dataset-generation dependencies are documented in the optional GPU section of requirements.txt. Keep those packages out of the production inference image.

API

• GET /health — liveness probe
• GET /v1/info — model + config metadata
• POST /v1/scan/image — scan an image (multipart/form-data, max 10 MB, JPEG/PNG/WebP)

See ARCHITECTURE.md for the full response schema.

Privacy

Visitor uploads are processed in-memory and never persisted. Only scan metadata (verdict, confidence, latency, model version) is recorded.

Documentation

• docs/plan.md — product roadmap with per-stage strengths, weaknesses, and fix paths. Written so a non-technical reader can follow the strategy, with deeper technical detail and a glossary inline.
• ARCHITECTURE.md — technical design of the detection pipeline + API contract
• docs/decisions.md — running decision log (good context if picking up this project later)
• NOTICES.md — third-party licensing record
• CLAUDE.md — project context (auto-loaded by Claude Code)
• scripts/dataset/README.md — dataset curation pipeline

License

Apache 2.0.