Compile the graph.
Inspect the computation.
Turn graph propagation into a reusable feature bank. Adjust its degree, perturb the inputs, and compare two independent paths through the same model.
A graph you can inspect
From edges to a feature bank
S = D⁻½(A + I)D⁻½Undirected graph · one self loopp(Σ BₖΘₖ + b)W + cPolynomial activation · linear readoutChebyshev recurrence: Bₖ = 2SBₖ₋₁ − Bₖ₋₂. The independent oracle constructs the full Tₖ(S) matrices. Both paths use identical inputs and weights.
Node scores
Raw logits are model scores, not calibrated probabilities. The largest score determines the prediction.
Measured in your browser
- Compile the bank
- —
- Evaluate compiled readout
- —
- Independent dense oracle
- —
- Mean absolute logit error
- —
One local run, including browser noise. These timings do not measure encrypted execution, a server, or a research benchmark.
Real CKKS · local prototype
Graph compilation is plaintext at the client. This prototype is not encrypted training, private-topology evaluation, remote key isolation, or a production security guarantee.
Move the graph work.
Keep the model explicit.
For a fixed graph operator, the client can compute Chebyshev-transformed features once. The readout then uses dense projections of those features, a polynomial activation, and a final classifier.
This laboratory checks the algebra with an independent direct oracle. It demonstrates numerical equivalence for the selected model; it does not establish accuracy on unseen data, conceal the default plaintext bank, or prove security.
Read the findings and proposed roadmapMethod reference: PrivNet-Krylov by Ehsan Hakimian and Chiman Soltanian ↗. This browser implementation was written independently; upstream source code is not bundled.