Flower Labs has developed a sovereignty-preserving opensource platform for building AI systems on distributed data. The aim is to help corporates and governments leverage siloed, private data at scale without compromising privacy, an ability which is especially valuable in regulated sectors like healthcare and finance, as well as edge-centric industries like telecom, energy and transportation, where data must remain onsite.
Users include JPMorgan, the UK’s National Health Service and Banking Circle, a fully licensed European bank providing financial infrastructure to payment companies, banks, and marketplaces.
“Today most AI relies on centralized public datasets,” says Daniel J. Beutel, Flower Labs’ Hamburg, Germany-based CEO and co-founder. Some 2000 trillion tokens of data are not public “an immense resource that is barely used today,” he says. “We believe unlocking access to orders of magnitude more private data will be game changing.”
AI thrives on data, but many verticals don’t have a lot of data in one place, says Beutel. “By allowing different jurisdictions or different companies to access a single analytics work flow, we can enable the use of AI in domains that really struggled to use it well beforehand.”
Flower Labs’ federated learning technology doesn’t require an exchange of data from client devices to global servers. Instead, the raw data on edge devices is used to train models or run analytics on private data locally, ensuring data privacy. For example, Banking Circle, which operates in both Europe and the U.S., needs to store its data locally due to regulations but it can use Flower Labs’ federated learning platform internally to leverage data on both sides of the Atlantic without moving it, and improve its training of AI models to detect money laundering, says Beutel.
In healthcare federated learning is invaluable because if each institution only uses traditional centralized training no one has enough data to develop breakthrough treatments, he says.
Governments can also benefit. “We believe that decentralized or federated AI is a key enabler for countries or regions to become more sovereign,” says Beutel. If a country or region is not very competitive in AI and uses federated learning to network together data across different organizations it can create an invaluable large, virtual decentralized data set which it fully controls and use it to train models that were impossible to train before, he says.
In September Flower launched SuperGrid, which aims to remove the biggest barrier to enterprise and government adoption of federated AI: the complexity of creating and managing federations. SuperGrid makes connecting distributed data sources and compute simple, without specialized setup or deep technical expertise, Beutel says. “We want to make it 10X easier to build these federations and collaborate on a global scale,” he says. “In the weeks and months to come we will be opening up to more and more users.”
With confidential computing on SuperGrid, untrusted parties can be kept outside the trust boundary, and workload correctness can be cryptographically verified through application-level attestation. These capabilities, grant competitors or nation states the ability to collaborate on mutually beneficial problems for the first time, he says.
Enterprise federated AI deployments need to satisfy a long list of security, compliance and operational requirements such as strict data protection, infrastructure scalability, high workload throughput, multi-tenant operations, and the ability to train the next generation of foundational models, They must also satisfy governance, audit, and compliance constraints that vary across regulated domains. Flower Labs’ architectural patterns and capabilities — strengthened further by SuperGrid – directly address these requirements, Beutel says.
Competitors include opensource communities such as OpenMined and Fed-BioMed, startups like Switzerland’s Tune Insight and the UK’s Flock.io, as well as federated offerings from companies like Google and Nvidia. Nvidia has added Flower support to its offering.
Beutel says Flower Labs’ platform is the most used framework in its space. In the last week alone, it says more than 75,000 models were trained on its platform.
The company, which is incorporated in Delaware and works remotely across Europe and the U.S., is currently focused mostly on the U.S. market. It has raised a total of $23.6 million. Mozilla is both an investor and a user of its technology.
This article is content that would normally only be available to subscribers. Become a subscriber to see what you have been missing
