Concrete ML FHE Demo ("SynthID Detector in the Browser")

Overview – Watermark Detection under FHE

This project implements detecting generative‑AI watermarks in fully homomorphic encryption (FHE) encrypted text. It adapts the principles behind DeepMind's SynthID watermarking scheme to text and executes the detector entirely on encrypted data. The client keeps all clear‑text locally; the server only ever sees ciphertexts.

At a glance:

Step	Who does it?	In clear‑text?	Purpose
Key generation	Browser (WASM)	✅	Produce client & server FHE keys
Tokenisation & encryption	Browser (WASM)	✅ → 🔒	Convert text → integer tokens → ciphertext
Watermark detection	Server	🔒	Compute SynthID score inside FHE
Result decryption	Browser (WASM)	🔒 → ✅	Show flag (= watermarked?) and score

---

Technology Stack

• tfhe‑rs – Rust implementation of the SynthID detection algorithm.
• concrete‑ml‑extensions‑wasm – WebAssembly bindings that enable client‑side keygen, encryption and decryption in the browser.
• Vanilla HTML/CSS/JS frontend – single‑page app with two web‑workers (keygen‑worker.js, encrypt‑worker.js) for heavy crypto.
• LLM Tokenizer (llama‑1b) – converts arbitrary text into integer IDs that the detector can process.

---

Data Flow in Detail

sequenceDiagram
    participant User🧑‍💻
    participant Browser🔐
    participant Rust Server🦀

    User🧑‍💻->>Browser🔐: 1️⃣ "Generate keys"
    Browser🔐->>Browser🔐: TFHE keygen (WASM)
    Browser🔐->>Rust Server🦀: 2️⃣ /handshake  (server_key_b64)
    Rust Server🦀->>Browser🔐: UID (uuid‑v4)

    User🧑‍💻->>Browser🔐: 3️⃣ Enter text
    Browser🔐->>Browser🔐: Tokenise ◁ HF tokenizer
    Browser🔐->>Browser🔐: Encrypt tokens (WASM)
    Browser🔐->>Rust Server🦀: 4️⃣ /detect (uid, ciphertext_b64)
    Rust Server🦀->>Rust Server🦀: FHE SynthID detect
    Rust Server🦀->>Browser🔐: Encrypted result_b64
    Browser🔐->>Browser🔐: Decrypt result (WASM)
    Browser🔐-->>User🧑‍💻: flag / score / g

Loading

Key points:

• The server currently stores the watermarking key in clear memory (SECRET_KEYS in detect_fhe_core.rs). Running crypting that key under FHE is possible with minor changes.
• Ciphertexts are serialised with bincode + Base64 for transport.
• A UUID returned by /handshake routes subsequent /detect calls to the right server key instance.

---

Demo Examples

Text fragment	Watermark score (≈)	Flag	Interpretation
Watermarked (see below)	0.6697	1	Highly likely to be generated by a watermarked model
Regular (see below)	0.5242	0	Looks like ordinary text

Watermarked text

The future of artificial intelligence for privacy is fully homomorphic encryption. While this still isn't fully accessible, many are exploring the potential of using it, which will help us in many different ways. This article will cover several examples of homomorphic encryption and see what its significance is for privacy.
Can Homomorphic Encryption be Applied to Privacy Issues?
Homomorphic encryption has been around for a very long time, especially in the areas of cryptography and mathematics. Some of it is for use in the military, while some of it in the private sector. Homomorphic encryption allows
Watermark score: 0.6697 (based on 110 tokens)

Regular text

The future of artificial intelligence for privacy is fully homomorphic encryption. Homomorphic encryption allows the secure transmission and execution of data without revealing the original data, preserving privacy. But how can we use future artificial intelligence to achieve greater privacy? Read on to find out.
Artificial intelligence is at an interesting place. It is becoming more powerful by the day, allowing it to run sophisticated analyses, even on extremely large datasets.
But this means that more and more personal data is being stored and used. The question is whether this will inevitably lead to more personal data
Watermark score: 0.5242 (based on 110 tokens)

---

Repository Structure (unchanged)

repo/
├─ server/            # Rust crate that exposes /handshake and /detect (Actix‑web)
│  ├─ src/
│  └─ tests/          # Rust unit & integration tests
├─ web/               # All static site files and web tooling live here
│  ├─ index.html
│  ├─ wasm-demo.js
│  ├─ concrete-ml-extensions-wasm/
│  ├─ package.json, pnpm-lock.yaml, node_modules/
│  ├─ playwright.config.ts, e2e/, __tests__/, test-results/
│  └─ ...
├─ docker/            # Dockerfile (+ docker-compose.yml optional)
└─ README.md          # This file

Quick‑start commands (Makefile powered)

Action	Command
Run backend & frontend (development)	make run-dev
Only start the Rust server	make run-server
Only serve the static web UI	make run-web
Build & run the server inside Docker	make run-docker
Spin up the full CI stack (Rust + Playwright)	make run-docker-compose
Execute Rust unit & integration tests	make test-rust
Execute browser end‑to‑end tests	make test-browser
Clean build artefacts & node_modules	make clean

All those targets live in the root‑level Makefile. Check it out to tweak ports, flags, or timeouts as needed.