On-device learning, illustrated live
0 raw examples leave the device
Privacy budget ε ≤ 10 per tenant and model.
A live demonstration of the technique Veil uses, not Veil's own results.
Concept demo — illustrates the design, no measured results.
What leaves the phone
Three phones train one shared model. Pick what leaves the phone and watch the path respond.
- Phone: its raw training examples stay on the phone.
- Path to the server: carries model updates.
- Path back: carries the improved model.
0 raw examples leave the device. Only model updates travel.
Illustrated flow, not measured traffic.
The trade-off, live
Three phones hold toy scores. This page averages them, adds Laplace noise set by your ε, and releases only the noisy average. Nothing leaves your browser.
| Phone | Scores kept on the phone |
|---|---|
| Phone 1 | 5 · 4 · 5 · 3 |
| Phone 2 | 2 · 3 · 2 · 4 |
| Phone 3 | 4 · 4 · 3 · 5 |
Veil caps real budgets at ε ≤ 10 per tenant and model; this slider stops at 10.
Expected wobble falls as ε rises. Drawn live from the toy's own math, not measured.
Toy truth 3.67 · released 3.67 · off by 0.00 at ε = 2. Expected wobble at this ε: ±0.17. Each draw spends one made-up unit; this demo reports no measured spend.
The honest numbers
- 0
- raw examples leave the device. The flow above never opens the raw path; only model updates travel.
- ε ≤ 10
- privacy budget per tenant and model. The ε slider stops at 10.
Everything else on this page is mechanism illustration, labelled where it appears: toy scores, made-up noise draws, and an illustrative budget meter.