Here's a nice magic trick: teach a model by watching a thousand people do a thing, without ever seeing what any of them did.
That's federated learning, roughly. And it's one of the most quietly important ideas in modern AI — because it lets us train useful models without hoovering up everyone's private data into a single server somewhere.
The dishwasher story again
Remember the dishwasher I talked about in the edge computing post? Imagine a thousand of them, in a thousand kitchens, each learning what "properly clean" looks like in its particular household. Each dishwasher improves, locally.
Now imagine the manufacturer wants to make all dishwashers smarter — to share the learning across the fleet, without ever seeing inside anyone's kitchen.
The way this usually works is: the cloud keeps a master model. Each device periodically calculates how the master model would need to change to better explain its local data — an update. Only the update gets sent to the cloud. The cloud averages thousands of updates from thousands of devices, applies them, and sends the improved master back out.
The raw data never leaves the device. The manufacturer learns from everyone, sees no one.
That's federated learning.
Why it matters
Privacy. This is the obvious one. A model can be trained on everyone's keyboard typing patterns, without anyone's typing ever being uploaded. Your phone's autocorrect uses this.
Regulation. GDPR, HIPAA, and similar regimes make it very hard to move certain kinds of data across organisational boundaries. Federated learning sidesteps the problem — nobody moves the data.
Edge cases that are actually edge cases. If you want a model to work well for rare-but-important situations (a specific medical condition, an unusual type of industrial fault), you need examples from many places. Federated learning lets you pool the learning without pooling the data.
Trust between organisations. Two hospitals, two banks, two competitors — people who can't share raw data can still collaborate on a model.
Where it's being used
- Google's Gboard predicts your next word using a model trained federally across millions of phones. Your typing stays on your phone.
- Apple's Siri improves its speech recognition federally for similar reasons.
- Medical research consortia are using it to train diagnostic models across hospitals that can't legally share patient data.
- Financial fraud detection between banks — a model that catches fraud patterns visible across institutions, without any institution seeing another's transactions.
The honest limitations
Federated learning isn't a silver bullet, and it has some real gotchas:
Privacy is not automatic. A model update, in the wrong hands, can sometimes leak information about the data it was trained on. Serious deployments pair federated learning with other techniques — differential privacy adds mathematical noise to the updates; secure aggregation encrypts them so only the average can be seen. If someone's selling you federated learning without either, be suspicious.
It's slower. Training converges more slowly than centralised training, because the updates are happening across a messy, heterogeneous fleet of devices.
Devices lie. If you're training federally across user devices, some of those devices will be malicious. Defending against poisoned updates is an active area of research.
Debugging is hell. When something goes wrong with a federated model, you can't just look at the data. You have to infer what's happening from the updates, or carefully designed metrics. Teams routinely underestimate the operational cost of this.
A pattern that's growing on me
A middle ground that's becoming more common: federated fine-tuning of a cloud-trained base model.
The base model is trained centrally on public or licensed data. Individual organisations then fine-tune it federally for their domain — picking up the nuances of their language, their users, their workflows — without sending any of their proprietary data back.
It's a neat division of labour. General knowledge in the cloud, specific knowledge at the edge.
If you're considering it
Ask yourself:
- Do we have a real data-residency or privacy problem, or are we just hoping the word "federated" will impress someone?
- Can we tolerate slower, messier training in exchange for not moving the data?
- Do we have the engineering maturity for distributed systems?
If yes, yes, and yes — federated learning might genuinely be the right shape for your problem. If not, you probably want centralised training with good access controls.
We've helped a few teams think this through. Come and talk to us if you'd like a second opinion.
