Flower Lab: Training AI on sensitive data

As AI grows increasingly data-hungry, we must ask ourselves how this can be reconciled with privacy. This question becomes even more urgent when AI starts entering fields where data protection is key.

To strike a balance between privacy and AI training, Hamburg-based Flower Lab created a decentralised AI training system. Built by Daniel Beutel (PhD, Cambridge), Taner Topal (industry CTO) and Nicholas Lane (Cambridge ML professor and former head of Samsung's AI lab in Cambridge), the system aims to make sensitive data usable without compromising the privacy of the individuals it belongs to.

Patient records, financial transactions, industrial sensor logs and personal device data cannot be pooled together; regulations limit the degree to which they can be used in order to safeguard the rights of citizens (especially in Europe). At the core of the solution is the principle of 'AI federation'. Multiple machines train simultaneously, but each one shares with the main model only the results, or 'weights', derived from the computations. These are then pooled together to allow the core AI to perform the necessary computations.

Concrete use cases go from co-training of illness prediction models without ever sharing patient records, banks collaborating on fraud detection models without exposing transaction data to competitors, and automotive companies training on in-vehicle sensor data that never leaves the car. Federated learning enables these organisations to develop effective models while complying with regulations.

In the technology sector, which prides itself on “moving fast and breaking things”, compliance with regulations is often an underappreciated part of the story. Nevertheless, new laws such as the EU AI Act will likely impose GDPR-like constraints on enterprises. For this reason, training and serving models in a federated manner will become increasingly critical over time.

Sources: Flower Lab / Crunchbase

Founders: Daniel Beutel, Taner Topal and Nicholas Lane