ArcInfer

Private sentiment classification powered by multi-party computation on Solana.

ArcInfer classifies text as positive or negative without ever exposing the input. The entire inference pipeline — from embedding to classification — runs on encrypted data using Arcium's MPC network.

Demo

ArcInfer.mp4

How It Works

flowchart LR
subgraph Browser
A[Text] --> B[Embed 384d] --> C[PCA 16d] --> D[Quantize Q16.16] --> E[Encrypt x25519]
end
subgraph Solana
E --> F[Submit 16 ciphertexts]
end
subgraph Arcium MPC
F --> G[Secret-share and classify]
end
G --> H[positive / negative]

Loading

1. Browser — Text is tokenized and embedded into a 384-dim vector using all-MiniLM-L6-v2 (ONNX/WASM), reduced to 16 dims via PCA, quantized to Q16.16 fixed-point, then encrypted with the MPC cluster's public key
2. Solana — A transaction carries the 16 ciphertexts on-chain and queues MPC computation
3. Arcium — MPC nodes secret-share the data and evaluate a neural network collaboratively — no single node ever sees the input
4. Solana — The classification result (positive / negative) is emitted via a callback event

Architecture

crates/
arcinfer-core/ Zero-dep Rust library: fixed-point math, NN layers, classifier
arcinfer-inference/ Client-side pipeline: tokenizer, ONNX embedding, PCA, quantization
arcinfer-pipeline/ Integration tests and weight loading utilities
programs/
arcinfer/ Solana/Anchor program: on-chain orchestration and MPC callbacks
encrypted-ixs/ Arcis MPC circuits: encrypted neural network evaluation
app/ Next.js frontend: wallet connection, encryption, progress tracking
training/ Python training pipeline (PyTorch → ONNX → PCA → Q16.16 weights)

Neural Network

Property	Value
Architecture	16 → 16 → 8 → 2 (fully connected)
Activation	x² (MPC-friendly, no comparisons needed)
Parameters	426
Precision	Q16.16 fixed-point (32-bit integers, 16 fractional bits)
Accuracy	80.2% on SST-2 sentiment benchmark

Prerequisites

• Rust (nightly, managed via rust-toolchain.toml)
• Solana CLI v1.18+
• Anchor v0.32+
• Arcium CLI
• Node.js v18+
• Yarn v1.22+

Build & Test

Rust crates

cargo test --workspace --exclude arcinfer --exclude arcis-arcinfer

Solana program + MPC circuits

Circuits are stored offchain on IPFS. Set their URLs before building:

export CLASSIFY_CIRCUIT_URL="<your-ipfs-gateway>/bafybeigjlp3cywmyd6xufivziaeedzfvlig6u2z6m4afsslaphmbblenzi"export CLASSIFY_REVEAL_CIRCUIT_URL="<your-ipfs-gateway>/bafybeidgrpe3wvpr3b3p46ipc4nmsjguwq3afe47o5nxqvvcxwqe23bwuu"
arcium build
arcium test

Any IPFS gateway works (e.g. https://ipfs.io/ipfs/, https://dweb.link/ipfs/, or a Pinata dedicated gateway).

Frontend

cd app
yarn install
yarn dev

The frontend needs model files and ONNX Runtime WASM binaries (not tracked in git):

# ONNX model for in-browser embedding
mkdir -p app/public/models
curl -L -o app/public/models/model.onnx \
"https://huggingface.co/Xenova/all-MiniLM-L6-v2/resolve/main/onnx/model.onnx"# ONNX Runtime WASM binaries
mkdir -p app/public/onnx
cd app/public/onnx
curl -LO "https://cdn.jsdelivr.net/npm/onnxruntime-web@1.21.0/dist/ort-wasm-simd-threaded.jsep.wasm"
curl -LO "https://cdn.jsdelivr.net/npm/onnxruntime-web@1.21.0/dist/ort-wasm-simd.jsep.wasm"

Devnet

Resource	Value
Program ID	2UEesrBiknFE3BoAh5BtZwbr5y2AFvWe2wksVi3MqeX9
Network	Solana Devnet
Circuits	Hosted on IPFS, fetched by Arcium nodes at runtime

Tech Stack

• Rust — Core math, NN inference, fixed-point arithmetic
• Solana / Anchor — On-chain program, transaction orchestration
• Arcium / Arcis — MPC circuit definition, encrypted computation
• TypeScript / Next.js — Frontend with browser-side ONNX inference
• ONNX Runtime (WASM) — In-browser sentence embedding (all-MiniLM-L6-v2)
• Tailwind CSS — UI styling

License

MIT