As you may already know, there is a lot of data out there and some of it may actually be quite useful. But privacy and security considerations often place severe limitations on how it can be used or analyzed. DataFleets promises a new approach whereby databases can be accessed and analyzed securely without the possibility of privacy breaches or abuses, and has raised an initial $ 4.5 million round to increase it.
To work with the data, you need to have access to it. If you are a bank, it means transactions and accounts; if you are a reseller, that means inventory and supply chains and so on. There are many usable insights and patterns buried in all of that data, and it is up to data scientists and their people to extract it.
But what if you I can not access data? After all, there are many industries where it is not recommended or even illegal to do so, such as in the healthcare sector. You can't exactly take medical records from an entire hospital, hand them over to a data analysis company, and say "sift through them and tell me if anything is good." These, like many other datasets, are too private or sensitive to be allowed anyone unlimited access. The slightest mistake - not to mention abuse - could have serious repercussions.
In recent years, some technologies have emerged that allow for something better, though: analyzing data without ever actually exposing it. It seems impossible, but there are computational techniques to allow data manipulation without the user ever actually having access to any of it. The most used is called homomorphic cryptography, which unfortunately produces a huge reduction in efficiency by orders of magnitude - and big data is about efficiency.
This is where DataFleets intervenes. He didn't reinvent homomorphic cryptography, but he somehow evaded it. It uses an approach called federated learning, where instead of bringing the data to the model, they bring the model to the data.
DataFleets integrates with both sides of a secure gap between a private database and the people who want to access that data, acting as a reliable agent to transfer information between them without ever revealing a single byte of actual raw data.
Here is an example. Suppose a pharmaceutical company wants to develop a machine learning model that examines a patient's history and predicts whether they will have side effects with a new drug. The private database of patient data of a medical research facility is the perfect thing to train him. But access is highly limited.
The pharmaceutical company analyst creates a machine learning training program and places it in DataFleets, which contracts with both them and the facility. DataFleets translates the model into its own proprietary runtime and distributes it to the servers where the medical data reside; within that sandbox environment, he transforms into a young ML agent, who is translated back into the analyst's preferred format or platform when finished. The analyst never sees the actual data, but he has all the advantages.
It's pretty simple, right? DataFleets acts as a kind of trusted messenger between platforms, undertaking analytics on behalf of others and never storing or transferring sensitive data.
Many people are looking into federated learning; the hard part is building the infrastructure for a broad enterprise-grade service. You need to cover a huge amount of use cases and accept a huge variety of languages, platforms and techniques, and of course do it in total security.
"We pride ourselves on business readiness, with policy management, identity access management and our SOC 2 certification pending," said DataFleets COO and co-founder Nick Elledge. "You can build anything on top of DataFleet and connect your tools, which banks and hospitals will say was not true of previous privacy software."
But once federated learning is set up, the benefits are suddenly huge. For example, one of the big problems in the fight against COVID-19 today is that hospitals, health authorities, and other organizations around the world have difficulty, despite their will, to securely share data related to the virus.
Everyone wants to share, but who sends who what, where is it kept and under whose authority and responsibility? With the old ways, it's a confusing mess. With homomorphic cryptography it is useful but slow. With federated learning, in theory, it's as easy as turning someone's access on or off.
Since data never leaves their "home," this approach is essentially anonymous and therefore highly compliant with regulations such as HIPAA and GDPR - another big plus. Elledge notes, "We are used by major healthcare institutions who recognize that HIPAA does not offer them sufficient protection when they make a dataset available to third parties."
Obviously there are less noble, but no less feasible examples in other sectors: wireless operators could make subscriber metadata available without selling out people; banks could sell consumer data without violating the privacy of anyone in particular; Bulky data sets such as videos can stay where they are instead of being duplicated and maintained at great expense.
The company's $ 4.5 million seed round is apparently proof of the confidence of a variety of investors (as summed up by Elledge): AME Cloud Ventures (Yahoo! Jerry Yang) and Morado Ventures, Lightspeed Venture Partners, Peterson Ventures, Mark Cuban, LG, Marty Chavez (chairman of Harvard's supervisory board), Stanford-StartX fund and three unicorn founders (Rappi, Quora, and Lucid).
With just 11 full-time employees, DataFleets appears to be doing a lot with very little, and the seed round should allow for rapid scalability and maturation of its flagship product. "We had to abandon or postpone the new customer demand to focus on our work with our lighthouse customers," Elledge said. They will hire engineers in the US and Europe to help launch their planned self-service product next year.
"We are moving from a data ownership economy to a data access economy, where information can be useful without transferring ownership," Elledge said. If his company's bet is successful, federated learning is likely to be a big part of this going forward.
