FSS is a prototype demonstrating a privacy-preserving semantic search system using Fully Homomorphic Encryption (FHE), specifically the CKKS scheme via Lattigo, combined with peer-to-peer networking and data distribution via Iroh. This project aims to match user queries with directory entries based on semantic similarity without revealing either the queries or the directory content.
A Taiwan CPR 2025 Project: https://community-privacy.github.io/
-
Directory Preparation:
- Each directory entry (e.g., user profiles, interests) is transformed into embeddings using Sentence-BERT embeddings.
- Embeddings are normalized and encrypted using CKKS via Lattigo.
- Encrypted embeddings are uploaded and stored securely on Iroh, a decentralized storage solution.
-
Query Processing:
- User inputs a plaintext query in the browser.
- Query is transformed into embeddings using Xenova's transformer model directly in-browser.
- Embeddings are encrypted client-side via WebAssembly-compiled Go code.
- Encrypted query embeddings are uploaded to Iroh.
-
Encrypted Semantic Matching:
- The Go backend retrieves the encrypted query embedding from Iroh.
- It performs homomorphic dot-product computations between the query and each directory embedding, producing encrypted similarity scores without decrypting data at any point.
-
Results:
- The encrypted similarity scores are returned to the client.
- Optionally, the client decrypts and ranks these scores locally.
Make sure you have Go 1.21+, Iroh CLI installed, and your browser supports WebAssembly.
Run this command to install Go dependencies:
go mod tidy- Populate
directory_entries.jsonwith your directory entries. - Generate normalized embeddings for each entry and save them in the
embeddings/folder.
Then, run:
go run prep_directory.goThis generates an embeddings_manifest.json after encrypting the embeddings and uploading them to Iroh.
Run this command to start the backend:
go run server.goYour server will start running at:
http://localhost:5000
- Open
index.htmlin your web browser. - Enter plaintext queries that will be automatically transformed, encrypted, and securely matched with directory entries.
- Frontend: HTML + WebAssembly
- Embeddings: Xenova transformers (browser-side Sentence-BERT)
- Encryption: CKKS via Lattigo v5
- Storage: Iroh decentralized storage
- Backend: Go + Gin web framework
- Multi-party functionality (currently implementation is single-key FHE)
- Matching only if both parties consent
- Integration of Zero-Knowledge proofs (e.g., Noir) for additional verification of skillset, credentials, etc
Many people to thank in the making of this project: Micah, Ayush, Vivek, Anka, Emma and the rest of the CPR team and friends