How a machine learning model collects data without compromising privacy?

(Subscribe to our Today’s Cache newsletter for a quick snapshot of top 5 tech stories. Click here to subscribe for free.)

Smart home devices like speakers and smart watches collect and share data with other devices and systems over the network. These Internet of Things (IoT) devices are equipped with sensors and software that store a user’s private information like body measurements and location.

This stored data is used by the device makers to improve their products and services.

An improvement in a machine learning (ML) model, called ‘federated learning’, is said to enable companies to develop new ways of collecting anonymous data without compromising their privacy, according to researchers at Missouri University of Science and Technology.

What is ‘federated learning’?

Federated learning is a ML method used to train an algorithm across multiple decentralised devices or servers holding data samples. It doesn’t exchange data with the devices, meaning there is no central dataset or server that stores the information.

Also read | U.N. decries police use of racial profiling derived from Big Data

Standard ML models require all data to be centralised in a single server. Implementation of federated learning eliminates the need for maintaining a storage hub.

The term was first introduced in a 2016 Google study titled ‘Communication-efficient learning of deep networks from decentralized data.’

Google emphasised mobile phones and tablets, stating that modern devices contain special features like speech recognition and image models that can store large amounts of data.

Since then, Google has used the technique is various products, including Gboard, which provides text and phrase suggestions to keyboard. Google had said the suggestions may be sent to its other services, excluding what was typed or spoken by the user.

How this works

Federated learning aims to train an algorithm, like deep neural networks, on multiple local datasets contained in local nodes, without explicitly exchanging data. The general principle involves simply exchanging parameters between these nodes. Parameters include number of federated learning rounds, total number of nodes, and learning rate.

Also read | MIT, Harvard team develop new model to improve AI-based decision making

The distinct advantage of the model is its ability to reduce privacy and security risks by limiting the attack surface to only the device, rather than the device and the cloud, Google stated in the study.

Federated learning is said to have application in healthcare, where hospitals and pharmaceutical companies can exchange data for treating diseases without sharing private clinical information.