What is Federated Learning?
In federated learning, the Models trained simultaneously on several machines. This is done decentrally and without any surrender or exchange of sensitive information. Thus, the corresponding data remain with the respective owners at all times. The central analysis model only receives the learning results and the parameters from the individual models.
The great advantage is that the learning effect created by incorporating information from different Training data is massively strengthened. Training is done on several devices in parallel and thus the accuracy of the model increases.
Federated learning can provide machine learning companies with the opportunity to get development of data-driven processes and services even with limited data. Federated learning has the potential to save costs and generate high added value.
How does Federated Learning of Cohorts (FLoC) work?
Federated learning of cohorts, FLoC for short, is part of the Google Privacy Sandbox initiative. It is a type of web tracking. The usage behaviour of all users can be directly evaluated by the browser itself and users are grouped into certain categories. Users then receive interest-based advertising.
The mode of operation goes via hashing, by generating cohort IDs within the browser. This is done using SimHash algorithms. The browser history can be encrypted using hash values. Privacy is protected because FLoC replaces third-party cookies and fingerprinting.
Nevertheless, targeted advertising for users is possible. Users are grouped according to the corresponding values and can then be targeted with advertising. Cohort IDs can be accessed via an API. These are newly created every week. Developers receive, for example, through TensorFlow a federated learning framework that enables computations for decentralised data and derive custom user types for advertisers.
Which framework is used to set up federal learning?
A Possibility is the well-known machine learning framework "Flower".. This framework was developed in 2020 and is particularly advantageous due to its wide distribution and high scalability. The infrastructure has proven to be very successful, especially due to the high user-friendliness of Flower.