Federated Learning
Instead of sending data to a central server, it remains where it is generated – on devices or local servers. The model is trained directly on-site, and only the model updates (e.g. weights and gradients) are shared. This enables companies and organisations to improve AI models without disclosing sensitive data or breaching data protection laws. Ideal for applications such as personalised voice assistants, medical diagnoses or smart IoT devices.
Development status
Scientific basis
Development and implementation
Initial field trials
Evaluation and optimisation
Ready for practical use
What problem do many companies face?
Companies often have vast amounts of data spread across different locations. The aim is to use this data to train a neural network efficiently, whilst ensuring that it does not fall into the wrong hands.
How does the tool help to solve the problem?
Federated learning is based on a small-scale server architecture. The federated server broadcasts the machine learning model to various devices, which then train the model locally using their respective data. The data is then sent back to the cloud, where the server aggregates a new model.
How is the tool used within the company?
In a medium-sized company, machines are to be monitored using anomaly detection. Either each machine could have its own ML model, in which case, however, the model would be unable to adapt to new environments. Or each machine could be trained on local data, which might not be sufficient to generate an optimal model. This is where federated learning comes into play: the machines are provided with a pre-trained model and then upload their data to the cloud, enabling a new model to be generated. This also allows new devices to be added and ensures that operations continue uninterrupted in the event of individual device failures.
What is the time commitment involved for the company in raising this objection?
The implementation of Federated Learning (FL) takes 3–12 months for planning, infrastructure, model development and data protection compliance. Once up and running, it typically requires a few hours to a few days’ work per month for training cycles, maintenance and optimisations. Training and efficient integration can reduce the workload in the long term. The exact time required varies depending on the use case and resources.
What added value does the tool bring to the company?
Federated learning enables collaborative learning. Data privacy is safeguarded; no data is sent to third parties. Thanks to its simplicity, federated learning is highly efficient.
Contact persons
