Machine Learning: The Synergy of Federated Learning and Data Protection

Machine Learning: How Federated Learning combines AI and data protection

Machine Learning: How Federated Learning combines AI and data protection

Merge models instead of data

From MRI to machine control

Privacy, partitioning, and aggregation

The most important federated learning frameworks at a glance

When developing machine learning models, one can definitely rely on the principle “A lot helps a lot”. The quality and quality of algorithms are inextricably linked to the amount of relevant data available.

Big tech companies are in the comfortable position of having enough data on their users and having secured the ability to use it for model training. The situation is different for classic companies, especially in the B2B environment. There, it is usually not the AI providers themselves who collect and collect data, but user companies – for example in mechanical engineering, where they come from the installed sensors.

There are two options for training models with this data: either the model provider trains a central model with all the data, or the users train decentralized and independent models directly on-site. In the first case, the user companies hand over the data to the manufacturer, who collects it centrally, trains an AI model on it, and then makes it available to the application companies. The model learns about all users here.

In principle, however, companies are reluctant to pass on their data because they fear that they could thereby disclose sensitive information such as production processes or intellectual property. The alternative is for each user company to create its own model. This solves the data protection and data security concerns, but the quality of the models is significantly worse than a model that has access to all data.

This is especially true when it comes to forecasting events such as machine malfunctions. Problematic courses form a strong minority here – these are so-called rare events. With a limited database, the training of a prognosis model is therefore considerably more difficult and reliable results are either not available at all or only appear with a long lead time.

More and more knowledge. The digital subscription for IT and technology. All exclusive tests, guides, & background information. One subscription for all magazines: c’t, iX, MIT Technology Review, Mac & i, Make, read c’t photography directly in the browser. No risk: first month free, then monthly from 9.95 €. Magazine subscribers read even cheaper! Start a FREE month. Try it now for FREE & continue reading right away! Already subscribed to heise+? Register and read. Register now and read the article immediately. To the start page.

Leave a Reply