DeepShield

AI armor against deepfakes and digital violence.

DeepShield is a browser-first toolkit for people facing deepfake abuse, morphed images, and non-consensual intimate imagery (NCII). It combines on-device computer vision, server-side ML scoring, reverse-image trace, encrypted evidence storage, legal PDF export, and trauma-informed support — in one web app.

Repository: github.com/Siddharthye/Deepshield

Live deployment

App	Role	URL
Frontend	UI, auth, client-side detection	deepshield-frontendd.vercel.app
Backend API	Hugging Face, LLM, trace APIs	deepshield-xi.vercel.app

The frontend proxies most /api/* routes to the backend via deepshield-frontend/vercel.json. Google OAuth routes stay on the frontend.

Features

Tool	Route	What it does
Scan	`/scan`	Image + video deepfake detection with heatmap and compare slider
Trace	`/trace`	Reverse image search (Google Lens) + manual URL logging
Report	`/report`	Court-ready PDF with scan data, trace log, and legal summary
Vault	`/vault`	PIN-protected AES-256 encrypted evidence storage (local)
Asha	`/asha`	Trauma-informed AI companion for rights guidance
Rights	`/rights`	Plain-language Indian cyber-law summaries
Community	`/community`	Anonymous local support feed with LLM moderation
Learn	`/learn`	Education hub — authentic vs suspicious examples, quizzes

Sign-in: Google OAuth or local email/password (stored in the browser only).

Languages: Full English and Hindi UI; 8-language selector with English fallback for other locales.

How detection works

Upload (browser)
    │
    ├─ face-api.js        → symmetry + face box
    ├─ Canvas artifacts   → edge / compression signals
    ├─ morphDetection.ts  → blend / seam heuristics
    │
    └─ POST /api/scan-image
           ├─ Hugging Face deepfake classifier
           └─ Sightengine (optional, parallel)
    │
    └─ riskScoring.ts → verdict + heatmap → optional LLM explanation

Important: The LLM does not decide if media is fake. It only explains detector outputs and helps with legal wording.

Privacy: Scans are sent to the API for scoring only — no server-side gallery of uploads. Vault and trace logs stay in the browser unless the user exports them.

Tech stack

Frontend (`deepshield-frontend/`)

Next.js 16 · React 19 · Tailwind CSS v4 · Framer Motion · GSAP · Three.js / R3F · Lenis · face-api · FFmpeg.wasm · pdf-lib · CryptoJS

Backend (`deepshield-backend/deepshield-backend/`)

Next.js Route Handlers · Hugging Face Inference · Sightengine · SerpAPI / Serper · in-memory rate limiting

Deployment

Two Vercel projects from this monorepo.

Project structure

deepshield-backend/                 ← Git root
├── deepshield-backend/             ← API app (deploy separately)
│   └── src/app/api/                ← scan, explain, asha-chat, trace, …
├── deepshield-frontend/            ← UI app (deploy separately)
│   ├── vercel.json                 ← proxies /api/* to backend
│   └── src/
│       ├── app/                    ← pages + Google OAuth routes
│       ├── components/
│       ├── context/
│       └── lib/                    ← riskScoring, i18n, pdf-generator, …
└── README.md

Local development

Prerequisites

Node.js 20+
npm

1. Backend API

cd deepshield-backend/deepshield-backend
cp ../.env.local.example .env.local
# Add HF_API_TOKEN at minimum
npm install
npm run dev

Runs on http://localhost:3000 by default (or the next free port).

2. Frontend

cd deepshield-frontend
cp .env.local.example .env.local
npm install
npm run dev

Set NEXT_PUBLIC_API_URL to your local backend URL, or use same-origin with rewrites if both run on the same host.

For local Google sign-in, add this redirect URI in Google Cloud Console:

http://localhost:3000/api/auth/google/callback

Confirm the live redirect URI after deploy:

GET https://YOUR-FRONTEND/api/auth/google/setup

Environment variables

Backend (`.env.local` in `deepshield-backend/deepshield-backend/`)

Variable	Required	Purpose
`HF_API_TOKEN`	Yes	Hugging Face inference + LLM
`HF_DEEPFAKE_MODEL_URL`	No	Default: `dima806/deepfake_vs_real_image_detection`
`HF_LLM_MODEL_URL`	No	Default: `meta-llama/Llama-3.1-8B-Instruct`
`SIGHTENGINE_API_USER` / `SECRET`	No	Boosts scan accuracy
`SERPAPI_API_KEY` or `SERPER_API_KEY`	No	Automatic reverse-image trace

See .env.local.example for the full list.

Frontend (`.env.local` in `deepshield-frontend/`)

Variable	Required	Purpose
`NEXT_PUBLIC_API_URL`	Yes	Use `same-origin` with `vercel.json` rewrites
`GOOGLE_CLIENT_ID`	For OAuth	Google OAuth Web client
`GOOGLE_CLIENT_SECRET`	For OAuth	Google OAuth secret
`GOOGLE_REDIRECT_URI`	No	Override if auto-detection is wrong

See deepshield-frontend/.env.local.example.

Never commit .env.local or secrets.

Deploy on Vercel

Backend project — root directory: deepshield-backend/deepshield-backend
Add HF token and optional Sightengine / SerpAPI keys.
Frontend project — root directory: deepshield-frontend
Set NEXT_PUBLIC_API_URL=same-origin, Google OAuth vars, and update vercel.json rewrite destinations to your backend URL.
Google Console — authorized redirect URI must match exactly:
https://YOUR-FRONTEND.vercel.app/api/auth/google/callback
Redeploy the frontend after any vercel.json or auth changes.

API routes (backend)

Route	Method	Purpose
`/api/scan-image`	POST	HF + Sightengine model scores
`/api/scan-video-frame`	POST	Per-frame video scoring
`/api/explain`	POST	LLM risk explanation
`/api/asha-chat`	POST	Streaming Asha chat
`/api/rights-explainer`	POST	Legal Q&A
`/api/reverse-trace`	POST	Google Lens lookup
`/api/trace-upload`	POST	Host image for trace
`/api/moderate-post`	POST	Community moderation
`/api/translate`	POST	UI string translation

Auth routes (/api/auth/google/*) run on the frontend only.

Crisis resources

DeepShield surfaces these contacts in-app:

Cybercrime helpline: 1930
Report portal: cybercrime.gov.in

Scripts

# Frontend
cd deepshield-frontend && npm run dev
cd deepshield-frontend && npm run build

# Backend
cd deepshield-backend/deepshield-backend && npm run dev
cd deepshield-backend/deepshield-backend && npm run build

License

Academic project. Check with the repository owner before commercial use.

Contributors

Siddharthye — github.com/Siddharthye/Deepshield

Navigation Menu

Repository navigation

About

Repository files navigation

DeepShield

Live deployment

Features

How detection works

Tech stack

Frontend (`deepshield-frontend/`)

Backend (`deepshield-backend/deepshield-backend/`)

Deployment

Project structure

Local development

Prerequisites

1. Backend API

2. Frontend

Environment variables

Backend (`.env.local` in `deepshield-backend/deepshield-backend/`)

Frontend (`.env.local` in `deepshield-frontend/`)

Deploy on Vercel

API routes (backend)

Crisis resources

Scripts

License

Contributors

About

Releases

Packages

Contributors

Languages

Name		Name	Last commit message	Last commit date
deepshield-backend		deepshield-backend
deepshield-frontend		deepshield-frontend
.env.local.example		.env.local.example
README.md		README.md

Latest commit

History

Folders and files

Repository files navigation

DeepShield

Live deployment

Features

How detection works

Tech stack

Frontend (deepshield-frontend/)

Backend (deepshield-backend/deepshield-backend/)

Deployment

Project structure

Local development

Prerequisites

1. Backend API

2. Frontend

Environment variables

Backend (.env.local in deepshield-backend/deepshield-backend/)

Frontend (.env.local in deepshield-frontend/)

Deploy on Vercel

API routes (backend)

Crisis resources

Scripts

License

Contributors

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

Frontend (`deepshield-frontend/`)

Backend (`deepshield-backend/deepshield-backend/`)

Backend (`.env.local` in `deepshield-backend/deepshield-backend/`)

Frontend (`.env.local` in `deepshield-frontend/`)