Este vídeo pertenece al curso Künstliche Intelligenz und Maschinelles Lernen in der Praxis de openHPI. ¿Quiere ver más?
An error occurred while loading the video player, or it takes a long time to initialize. You can try clearing your browser cache. Please try again later and contact the helpdesk if the problem persists.
Scroll to current position
- 00:00Welcome to the excursion Topic Federated Learning.
- 00:04As in the others Excursions under the title Hot Topic,
- 00:08we will also be using Federated Learning This topic is relatively new and innovative.
- 00:14But first of all, the starting point: How does machine learning traditionally work?
- 00:22So let's imagine this scenario.
- 00:27We have a scenario where four actors are involved hold a portion of the data set.
- 00:34In order to do machine learning, you first have to do the various Merge partial data sets at a central location.
- 00:42At this central point, the Exercise the model with different parts of the data set.
- 00:50So what you have to notice is that you have to train Machine learning usually involves data in a central location
- 00:56and a single, accessible unit and access to the data.
- 01:03If the model is then trained on all data, then the central model can be used by the individual actors
- 01:10the central model can be individual actors.
- 01:15Or the individual actors make inquiries to the model For example, only the result is played back.
- 01:25But what do you do when the individual entities cannot share their data with the central unit?
- 01:32One possibility, of course, would be that everyone Actor trained only on his own data.
- 01:39But it can happen here that too few Training data for machine learning are available
- 01:44or with a larger Record could make significantly better predictions.
- 01:51In which scenarios and in But what are the applications?
- 01:57We will briefly consider this in the following.
- 02:00Here are a few Sample applications.
- 02:05Many prominent examples will for example, in our smartphones.
- 02:12We always want clever proposals for the have the next word in a text input on your smartphone.
- 02:19However, of course our typed Messages do not leave our smartphone.
- 02:26Another known example It's from the public health.
- 02:32We want patients provide models that are as intelligent as possible,
- 02:36e.g. for cancer detection, or to predict hospital stays.
- 02:41without the sensitive personal Data in quotation marks must leave the hospital.
- 02:49Here comes a Federated Learning to or can be used.
- 02:54Federated Learning enables machine learning across organizational and entity boundaries,
- 03:00Without disclosing sensitive data.
- 03:05The benefits of this process are having a better model. which can be trained on more or different data
- 03:12without having to share data.
- 03:17A machine learning diagnostic model could learn from cases involving several hospitals,
- 03:22without patients and Patients must fear for their data.
- 03:28However, Federated Machine Learning not the only method in this field.
- 03:33The field can be roughly under Privacy preserving machine learning.
- 03:39Another approach would, of course, be to reduce the anonymize and provide sensitive information
- 03:46conceal or remove.
- 03:50As a further approach, which is also relatively new with many new research results,
- 03:56is machine learning on encrypted data.
- 04:00The data is stored only encrypted at a central Entity that then tries to make predictions
- 04:07on just that encrypted data.
- 04:11Homomorphic Encryption is a This is exactly what a promising approach is for.
- 04:18Homomorphic encryption is a new encryption method. to enable calculation with encrypted data
- 04:25as if they would be unencrypted.
- 04:30The central entity has no Ability to decrypt the data.
- 04:37This is made possible because the inherent structure is maintained, despite encryption.
- 04:44And the approach, of course, is Federated Machine Learning.
- 04:49Federated Machine Learning sets in contrast to the other This does not presuppose that the data must be centrally available.
- 04:57But how does it work exactly?
- 05:02We shall now Take a closer look.
- 05:07First of all, each node trains or any actor on their own data
- 05:12And it builds its own local model.
- 05:20This locally trained model is then subsequently sent to a central body,
- 05:25where it is aggregated with other local models.
- 05:29What's important is that we just have the model and the lessons learned. parameter, but not the original data.
- 05:39Although there are ways to get from the learned parameters draw conclusions about the original data;
- 05:46But there are, of course, appropriate Countermeasures in federated learning.
- 05:52For the central model if then the weights and a central consolidated model.
- 06:00The central model, in turn, updated local models.
- 06:06The different local models then update over time We're going to go back to the global model.
- 06:13Then the cycle starts again and the central model aggregates then re-updates the local models.
- 06:23The central importance of Federated Learning, that we use the models or the model updates
- 06:29of the local models that are sent to the central entity , do not implicitly disclose data.
- 06:39That's it too our short excursion
- 06:43Federated Learning and its applications.
To enable the transcript, please select a language in the video player settings menu.