Can you run a GNN without anyone seeing the data? Journal of our experiments on privacy-preserving GNN inference on microcontrollers
I have been keenly interested in how to make Graph Neural Networks (GNN) run securely on resource-constrained devices. The question driving this work is deceptively simple: can three microcontrollers jointly compute GNN inference on traffic data such that no single device ever sees the plaintext , not the input features, not the model weights, not the intermediate activations? This post is about…
The article investigates the feasibility of running Graph Neural Networks (GNN) on microcontrollers without exposing sensitive traffic data. This is particularly important in scenarios where multiple parties need to collaboratively optimize traffic signals at shared borders, as raw traffic data can reveal movement patterns, congestion bottlenecks, and individual vehicle trajectories.
The authors propose a solution using Replicated Secret Sharing (RSS) on three microcontrollers, ensuring no single device ever sees the plaintext data.
Key to this approach is recognizing that the graph topology (adjacency matrix) is public knowledge, while only the node features and model weights need protection. This allows the message-passing step (A · H) to be computed locally without communication, drastically reducing the computational burden. The only step that requires secure matrix multiplication is H · W, which is handled using RSS.
The researchers use a 3-party semi-honest setting for RSS, where each party holds two shares of the secret. Secure addition can be performed locally without communication, while secure multiplication incurs a single round of communication using a PRF-based resharing protocol. The authors train a 2-layer GCN on a synthetic traffic grid, using a polynomial activation function to minimize computational complexity in MPC.
The fixed-point quantization preserves inference fidelity, with minimal error even after converting to Q16 format. The entire pipeline—from training to inference verification—is tested on ESP32-S3 microcontrollers, demonstrating that secure GNN inference is achievable on resource-constrained devices while preserving privacy.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — it may contain errors, so check the original before relying on it.
