Patient files, account histories, production logs: the law or the contract says they stay. Each site alone has too few cases to learn from. A central training set is off the table.
Patient files, account histories, production logs. The law, the contract or the board says they stay where they are. A central training set is off the table.
A data sharing agreement across three organisations takes a year and a legal team. Most projects do not survive the wait.
One hospital has a few hundred cases of the condition. Five hospitals together would have enough to learn from. Apart, none of them can.
Each site trains the model on its own records, inside its own perimeter. Only the model updates travel, combined in a way that reveals none of them. The result is one model that learned from all the sites, and no record ever moved.
Three parts: training at each site, combining the updates securely, and sending the result back.
The model runs at each participant, on their machine.
The coordinator combines what it cannot read.
Everyone gets the improved model back.
Four phases, from the participants around the table to a model at every site.
Participants, data, rules, the question.
Clients at each site, coordinator in the middle.
Rounds, evaluation, repeat.
The model at every site, monitored.
A model trained on data that could never be pooled, with the compliance argument written into the design.
No data sharing agreement, because no data is shared. Legal reviews a protocol instead of a transfer.
Measured at each site against a model trained on its own data alone. The gain from the other sites is a number, per site.
A site can see what it contributed, pause, or leave. Its records never left its building, so there is nothing to recall.
| Side by side | Central training | Federated learning |
|---|---|---|
| Where the data goes | To a central store | Nowhere |
| Privacy | Depends on the store's security | Records never leave; updates combined unseen |
| Compliance | Data sharing agreements, transfers, residency | A protocol review |
| Data the model sees | What could be transferred | Every participating site |
| Infrastructure | One large training environment | A client per site, a coordinator in the middle |
A way to train one model on data held in several places without moving the data. Each place trains on its own records and sends back only the model update. A coordinator combines the updates into a better model and sends it back. Repeat until the model stops improving. No record leaves the place it lives.
First, raw data never leaves a site. Second, the updates are combined with secure aggregation, so the coordinator sees only the sum, never one site's contribution. Where a stricter guarantee is needed, noise is added to the updates. The protocol and the guarantee are written down and reviewed by each participant before a round runs.
Health, where patient records stay in each hospital. Financial services, where account data stays in each institution. Manufacturing groups with plants in several jurisdictions. Any case where several holders of similar data want one model and none of them can share the data.
Usually close, and always better than what any single site can train alone, which is the comparison that matters when pooling is not an option. We measure it at each site against the site-only baseline and report the numbers per site.
We are a small team of senior specialists. We pick the right model and the right layer, and we build the least machinery that does the job. You get a call with an engineer, not a sales deck.
