INSPIRATION Today, everyone wants to use powerful AI like Google Gemini or ChatGPT. But when companies or individuals upload sensitive documents like salaries, resumes, financial audits, or passwords, that private data is sent directly to cloud servers, risking leaks and unauthorized training. We asked a simple question: Can you ask AI any question about your confidential documents without the AI company ever seeing your private secrets? This is why we built BlindLLM.

WHAT IT DOES BlindLLM acts as a smart privacy shield between you and cloud AI:

  1. Automatic Local Masking: You upload a document (PDF, Word, or Text file) or ask a question. BlindLLM automatically detects sensitive data (names, phone numbers, emails, salaries, credentials) right in your browser.
  2. Safe Placeholder Tokens: It replaces secrets with harmless tags like [PERSON_1], [FINANCIAL_1], or [CONTACT_1].
  3. Secret Sharing Math: The encryption key that locks your secrets is split into 4 mathematical pieces using Shamirs Secret Sharing. No cloud provider ever receives the key.
  4. Cloud AI Answers Safely: Google Gemini reads the masked text, understands the context, and answers the question accurately using the tags.
  5. Instant Local Unmasking: When the answer returns to your computer, BlindLLM swaps the real names and numbers back in. You see the complete, natural response, while Google saw zero secrets.

HOW THE SECURITY WORKS

  1. 3-of-4 Shamirs Secret Sharing: The master encryption key is split into 4 pieces. It mathematically requires at least 3 pieces to unlock. An attacker or AI with fewer pieces learns 0 bits of information.
  2. Feldmans Verifiable Secret Sharing: Allows client nodes to mathematically verify their shares without decrypting them.
  3. Zero Key Bits Leaked: The cloud LLM only ever receives placeholder tokens.
  4. Active Memory Zeroize: As soon as your answer is displayed, temporary memory holding private keys is wiped clean to prevent memory leaks.

HOW WE BUILT IT

  1. Frontend: Built with React 19, TypeScript, and Tailwind CSS, featuring a live side-by-side view (What You See vs What Cloud AI Sees) and a real-time Cryptographic Telemetry dashboard.
  2. Backend: Built with Python and FastAPI, with local document extractors and 8 automated cryptographic test suites.
  3. Cloud AI: Powered by Google Gemini API (gemini-flash-lite-latest) for fast, context-aware reasoning.

CHALLENGES WE FACED

  1. Getting AI to Understand Placeholders: Prompting the LLM so it keeps placeholder tags intact rather than ignoring or changing them.
  2. Zero-Leakage Guarantee: Ensuring that neither names nor encryption keys ever cross the network wire.
  3. Seamless Local Unmasking: Reconstructing the secret key and restoring text in under 2 seconds.

ACCOMPLISHMENTS WE ARE PROUD OF

  1. Mathematically proven 0 key bits leaked to the cloud AI provider.
  2. Fast, live end-to-end response time in about 1.5 seconds.
  3. Full-stack working application deployed live on the web.
  4. Real-time side-by-side transparency view showing exact wire traffic.

WHAT WE LEARNED

  1. How threshold cryptography and secret sharing can be applied to modern AI inference pipelines.
  2. How to combine mathematical security with an intuitive, clean user interface.

WHAT IS NEXT

  1. Adding browser extension support to protect ChatGPT and Gemini web chats automatically.
  2. Supporting enterprise team vaults with hardware-based key storage.]=

Built With

  • aes-256-gcm
  • cryptography
  • elliptic-curves
  • fastapi
  • feldman-vss
  • finite-fields
  • google-gemini-api
  • pypdf
  • pytest
  • python
  • react
  • shamirs-secret-sharing
  • tailwindcss
  • typescript
  • vercel
  • vite
  • zero-knowledge
Share this project:

Updates

Submission history