Google built new technology to give consumers privacy by doing more with less data over time, with a focus on data processing on-device to make features faster and data more accessible for every type of application, from search to Maps.
Sundar Pichai, Google CEO, refers to that technology as on-device learning. He calls it Federated Learning -- a new approach to machine learning, introduced at Google’s developer I/O conference. It relies on the same artificial intelligence used to develop Google products, and now the company is applying it to user privacy.
Federated Learning does not collect raw data from consumer devices. Rather than sending data to the cloud, Google developers flipped the model and shipped the machine learning model directly on the device, Pichai said during the I/O conference.
Each phone computes an update to the global model. Only those updates, not the data, are aggregated and updated to improve the global model. Then the updated global model is sent back to the phones.
Google’s keyboard Gboard, using only on-device learning technology, would not allow it to update its model until many people typed the new words many times. Federated Learning allow Gboard to learn new words like “BTS” or “YOLO” after thousands of people use these new words without allowing Google to ever see anything people type.
Gboard is using Federated Learning to make next-word predictions across “tens of millions of devices,” Pichai said.
Pichai also introduced incognito in Google Maps, so the places searched in Maps will not link to a user’s Google account. He also talked about safe browsing to protect 4 billion devices, Tensorflow technology that prevents Phishing in Gmail, and a two-step verification process now available to all Android phones running version 7.0 or higher.