Collecting and sharing the data needed to train models that can offer valuable information through AI and machine learning poses significant challenges, both as a result of the enormous amount of quality information needed and the obligation to guarantee the security of this information. Data owners—government organizations, private companies, and research institutions—must ensure data privacy, security, and sovereignty, recognizing that sharing certain kinds of information can pose significant risks and that, in most cases, there are legal or political restrictions that limit data sharing.

As Pablo González, an artificial intelligence specialist at GMV, explained at his TED talk at the 1st Andalusian Artificial Intelligence Congress, organized by the government of Andalusia: “We’ve developed uTile, a tool based on PET (Privacy-Enhancing Technologies), that uses a federated learning approach applicable to any activity sector to solve the issue of how to work with large amounts of data while guaranteeing the privacy and governance of these data.”

Within this framework, the process of training the model is shared by model developers and data owners. Each party trains a portion of the model in their local environment, without needing to share the raw data directly. The parties then work together to pool their contributions and build a complete and accurate artificial intelligence model.

Earth observation success story.

González shared the success story behind the development of GMV’s uTile tool: Earth observation applying digital technology, in response to a European Space Agency challenge. As he explained, federated learning provides an effective solution in terms of preserving the privacy and security of data while they are being used to train advanced models. In this case, GMV's tool makes it possible to analyze how harvests are progressing, the risk of fire or floods, or the state of roads, by making it possible for software developers and owners of satellite photos to work together without having to share their data. Each one trains their part of the software on their own computer, and then they put the pieces together to have a complete program. It’s a win-win situation in which no one sees or sends the other party’s data.