Training with data kept on phones saves communication, not privacy
A vendor's pitch says "the model is jointly trained on users' phones, raw data never leaves the device" — behind it is most likely this paper's approach: devices train locally for multiple rounds, and the server only averages model parameters.
Long-term reading
A vendor's pitch says "the model is jointly trained on users' phones, raw data never leaves the device" — behind it is most likely this paper's approach: devices train locally for multiple rounds, and the server only averages model parameters. Communication rounds are ten to a hundred times fewer than synchronizing gradients every step, and it can train even when data is unevenly distributed across participants.
Next time you see "data never leaves the premises" training claims, first clarify what it actually saves: communication cost, not a privacy guarantee. Cross-device training went from infeasible to viable through this trade-off — train more rounds locally, transmit fewer times.
If the evidence comes only from simulated environments and proxy datasets, don't use it to infer performance on real phones. If you care about privacy attacks or malicious updates, the paper doesn't address them — don't cite its conclusions as backing. And in data-center settings where data is centralized and identically distributed, it was never needed in the first place.
《Communication-Efficient Learning of Deep Networks from Decentralized Data》(2016) | Next review 2027-09-20