Federated Learning Could Be the Next Big Thing for Data Privacy

“Federated learning.”
What it sounds like: a government-mandated curriculum.
What it is: an emerging, privacy-focused form of machine learning.

Google introduced the technique in 2017 to test on Android keyboard suggestions. Using it, smartphones could locally store and process info about which suggestions a user chose and surrounding context. That helped improve the keyboard algorithm without sacrificing as much user privacy.

• The traditional learning method was to collect user data, upload it to the cloud, then run an ML model. That meant sensitive user data was all stored in one place. (Hello, hackers.)

Hype cycle: In H1 2020, researchers published more than 1,000 papers on federated learning—compared to just 180 total papers in 2018. Google searches for the term look like 📈.

The technique is especially useful for sensitive data collected by apps and connected devices, says Ashwin Machanavajjhala, an associate professor and privacy researcher at Duke University.

• Federated learning allows your iPhone to wake to your “Hey Siri” but not your friend’s. Fitness and health apps are another likely use case. So is vehicle telemetry, where sensors collect data on driver/vehicle performance.

Looking ahead: As more and more sensitive info is collected about us, we’ll likely see federated learning deployed to try and stem the risks.