Apppublic

ofc01/janrakshak-api

sourceHugging Faceupdated 4mo agoView on Hugging Face

0likes

App README

🛡️ JanRakshak Vision — Backend API (v7)

AI Deepfake Detection Backend | JanRakshak Vision | Tradition Hacks 2026 Team: Anonymous Group | Leader: Kushal Soni

![HuggingFace Space](https://ofc01-janrakshak-api.hf.space) ![FastAPI](https://fastapi.tiangolo.com/) ![Python](https://python.org)

🌐 Live Deployment

Service	URL
API Base	https://ofc01-janrakshak-api.hf.space
Swagger UI	https://ofc01-janrakshak-api.hf.space/docs
Frontend	https://janrakshak-frontend.vercel.app

📡 API Endpoints

`GET /`

Health check and basic info.

json

{
  "status": "JanRakshak Vision API ✅",
  "version": "1.0.0",
  "team": "Anonymous Group",
  "leader": "Kushal Soni",
  "competition": "Tradition Hacks 2026",
  "limits": {
    "image_max": "50MB",
    "video_max": "100MB",
    "rate_image": "10 requests/minute per IP",
    "rate_video": "5 requests/minute per IP"
  }
}

`POST /analyze/image`

Upload any image (JPG, PNG, WEBP, GIF, BMP — max 50MB)

Response:

json

{
  "verdict": "FAKE",
  "confidence": 88,
  "fake_score": 0.8842,
  "real_score": 0.1158,
  "explanation": {
    "en": "This image shows 88% signs of AI generation...",
    "hi": "इस तस्वीर में 88% AI निर्माण के संकेत हैं...",
    "bn": "এই ছবিতে 88% AI তৈরির লক্ষণ রয়েছে..."
  },
  "file_type": "image",
  "file_name": "photo.jpg",
  "frames_analyzed": null,
  "model_votes": [
    { "name": "detector_v2", "fake_score": 0.95, "verdict": "FAKE", "confidence": 95 },
    { "name": "sdxl",        "fake_score": 0.82, "verdict": "FAKE", "confidence": 82 },
    { "name": "general",     "fake_score": 0.85, "verdict": "FAKE", "confidence": 85 }
  ]
}

`POST /analyze/video`

Upload any video (MP4, AVI, MOV, MKV — max 100MB)

Same response format + frames_analyzed count + frame_breakdown array (per-frame verdicts), plus avg_fake_score and real_score averaged across frames.

🤖 The "3-Brain" AI Ensemble (v7)

Three models run in parallel, combined with a custom Smart Heuristic Engine:

Model	HuggingFace ID	Weight	Specialty
detector_v2 (Gen-AI)	`haywoodsloan/ai-image-detector-deploy`	35%	Generative AI (Midjourney, DALL-E, SDXL)
sdxl (Texture)	`Organika/sdxl-detector`	35%	Pixel noise, texture artifacts
general (Composites)	`umm-maybe/AI-image-detector`	30%	Composite edits, face-swaps, Photoshop

Smart Heuristic Logic:

—High-Confidence Boost: If any single model >85% FAKE, final score is boosted.
—Composite Floor: If Brain 3 (Composites) ≥70%, final verdict floored to SUSPICIOUS.
—Screenshot Filter: Detects UI screenshots (low color count + low variance) and dampens the fake score by 60% to prevent false positives.
—Thresholds: FAKE ≥ 50%, SUSPICIOUS ≥ 30%, REAL < 30%

🏗️ Architecture

React Frontend (Vercel)
        │
        │ HTTPS POST /analyze/image
        ▼
FastAPI Backend (HuggingFace Space — Docker)
        │
        ├── Brain 1: haywoodsloan (35%)
        ├── Brain 2: Organika (35%)
        └── Brain 3: umm-maybe (30%)
                │
                └── Smart Heuristic Engine (Python)
                        │
                        └── FAKE / SUSPICIOUS / REAL

🚀 Local Development

bash

# Install dependencies
pip install -r requirements.txt

# Run locally
uvicorn app:app --reload --port 8000

# Test API
curl -X POST "http://localhost:8000/analyze/image" \
  -F "file=@test_image.jpg"

📁 Project Structure

janrakshak-backend/
├── app.py              # FastAPI routes, Rate Limiting (slowapi), CORS, lifespan
├── models.py           # 3-model ensemble inference & heuristic engine
├── utils.py            # PIL helpers, video frame extraction, memory mgmt
├── requirements.txt    # Python dependencies
└── Dockerfile          # HuggingFace Space Docker config

🔒 Privacy & Zero-Storage

—Zero storage — files analyzed in RAM, never written to disk (except temp video files, instantly deleted via os.unlink).
—Memory Management — explicitly calls del and gc.collect() after every inference to free memory instantly.
—No logging — image contents, user IPs, and hashes are never stored.

<div align="center"> Built for Tradition Hacks 2026 | Leader: Kushal Soni | Anonymous Group </div>