Google tool lets any AI app learn without taking all your data

TensorFlow Federated limits the impact of AI on your privacy.

Laura Hautala
Laura Hautala
Laura Hautala Former Senior Writer
Laura wrote about e-commerce and Amazon, and she occasionally covered cool science topics. Previously, she broke down cybersecurity and privacy issues for CNET readers. Laura is based in Tacoma, Washington, and was into sourdough before the pandemic.
Expertise E-commerce, Amazon, earned wage access, online marketplaces, direct to consumer, unions, labor and employment, supply chain, cybersecurity, privacy, stalkerware, hacking. Credentials
  • 2022 Eddie Award for a single article in consumer technology
Laura Hautala
3 min read
A circuit-board like design forms a profile of a human brain with ones and zeroes in the background.
Getty Images

A new computing tool developed by Google will let developers build AI-powered apps that respect your privacy.

Google on Wednesday released TensorFlow Federated, open-source software that incorporates federated learning, an AI training system. It works by using data that's spread out across a lot of devices, such as smartphones and tablets , to teach itself new tricks. But rather than send the data back to a central server for study, it learns on your phone or tablet itself and sends only the lesson back to the app maker.

Federated learning runs "part of the machine learning algorithm right next to where the data is on the device," Alex Ingerman, a product manager at Google Research, said in an interview. The algorithm applies what it already knows to the data on your phone, such as suggesting replies to emails, and creates a summary of what it learned in the process to send back.


TensorFlow Federated adds an important, new privacy-sensitive ability to the artificial intelligence revolution taking hold of the computing industry. AI promises to change the way we work and live, letting machines learn enough that they can complete tasks that currently require people. For example, if you and a bunch of other people add "side-eye" to your texting app's dictionary, the app could figure out the usage and incorporate that into its standard dictionary by itself.

To get good at these tasks, machines need to see a huge amounts of data, which worries people concerned about privacy . Federated learning helps soothe those worries.

Google has led the AI charge, using the technology for tasks like translating written languages spotted in photographs or suggesting responses to emails. TensorFlow Federated is already built into some Google apps, such as the Gboard keyboard for Android phones and iPhones, where it analyzes typing patterns in order to offer suggested shortcuts. Now that it's freely shared open-source software, TensorFlow Federated can help other developers with AI projects without requiring them to start from scratch.

Predicting what you want with AI is a major part of Google's business plans. Scott Huffman, Google's vice president of engineering, said in January that AI will help Google Home Assistant engage in basic conversation within five years. Eventually it will interpret your mood and remember the details of previous conversations. Google announced federated learning for its own apps in 2017. 

Federated learning builds on Google's TensorFlow system, a machine learning system that tech companies and academics use in their products and projects for free. For example, TensorFlow is part of Google's efforts with Stanford University researchers to explore potential new drugs. What's more, companies such as Twitter, Coca-Cola and Intel use the platform.

Apple addresses privacy in iOS apps that use machine learning through a process called differential privacy, which it started touting in 2016. Developers can create iOS apps that run machine learning locally on users' phones and tablets without extracting any raw data.

Ingerman said Google hopes academics will produce research using TensorFlow Federated that furthers everyone's understanding of the technology and its uses.

"Our aim is to give back to the research community and enable development in the field," Ingerman said.