Artificial intelligence has to learn from some amount of information in order to make decisions later. However, this information sometimes needs to be gathered somewhere specific in order for the intelligence to get to it. Federated learning provides a possible alternative to this scenario by allowing multiple entities to collaborate on training an entity without sharing all of the information they have.
Introduction
According to IBM, federated learning is a decentralized process during which a shared model is trained by many participants. These participants get the shared model to their individual data and train it on it. After that, they send an update of the changes to the shared model to a coordinating server, which combines these updates. The process is then repeated until the model achieves the desired result.
Therefore, federated learning is different from simply collecting the records of all the participants and training the model on this collective dataset. The shared model is updated in accordance with the data it is being trained on without this data being collected and combined.
## Federated Learning Use Case
IBM gives an example of multiple factories trying to train a model to predict equipment failures. There is a desire to share this information between the factories, but each one is hesitant to share its own records. Federated learning would allow the factories to install the shared model on their servers, update it according to their data and send the updates to a coordinator, which would combine them to form an updated version of the shared model.
This example shows how collaboration can take place even when entities do not share their data directly. The decision to participate in the collaboration depends on the value of the data to the participating entities and the trust they have in the other entities.
The use case also shows that although federated learning allows entities to keep their data private, this does not mean that all privacy concerns should be disregarded. IBM notes that a federated-learning system has various other privacy and security implications. Some of these, like differential privacy and secure multi-party computation, are discussed in the article, while others depend on the implementation, which should be audited by third parties. Additionally, participants should make sure that the shared model is actually beneficial to them and that they are not just training it for someone else.
Conclusion
Federated learning can allow entities to cooperate on creating an AI model without sharing their data with each other. Therefore, it can be beneficial in many situations where data privacy is an issue, but requires careful implementation to avoid other issues. The main advantage of federated learning is that it makes cooperation possible without compromising on data privacy.
Reference
IBM. “ What Is Federated Learning? ” Web.