Accountable Federated Machine Learning for public administration services in Germany
The AFML project at the Center for AI involves the development of a prototype that enables models to be trained for the classification of ideas submitted by citizens by means of federated machine learning (FML). Despite the decentralized approach, the system will be designed to ensure a traceable and verifiable process while guaranteeing that the results that are generated adhere to criteria such as data protection, security and accuracy. The concept is based on data and models, in the context of civic participation, that will be used to automatically group various input from citizens according to the subject or issue.
The plan is to utilize and further train these models beyond the boundaries of the city, but without directly exchanging information between the participants by locally training the respective model and transferring only the resulting changes to an aggregated model. The underlying principle here is “share knowledge, not data”. With the AFML approach, individual steps and local training iterations are logged in order to be able to verify the results and detect manipulation or errors.
Federated Machine Learning (FML) is a highly-promising and exciting field of research in the area of machine learning that makes it possible to locally train decentralized models and make them available to the various participants for use on a wide basis. The idea is also to be able to train models based on large volumes of data in areas where the exchange of data is not possible.
The fortiss research activities are focused on the area of “secure & privacy preserving machine learning”. In the AFML project, researchers will enhance the concept with a verifiability component so that responsibilities can be defined, results verified and manipulation or errors in individual process steps detected despite the decentralized nature of the FML approach. This enhancement of the current research creates trust in the use of decentralized machine learning models.