Formalizing federated learning and differential privacy for GIS systems in IIIf

GIS systems, like Google maps or ArcGIS, are a ubiquitous central application but are highly privacy critical. In many GIS systems, inputs from various and diverse sensors potentially expose private information. However, in particular, in the context of public safety, the sensor inputs are crucial to provide timely information for example for weather related warning systems. The notion of Differential Privacy (DP) has become an ad hoc standard, most notably adopted for the US census. Federated Learning (FL) facilitates machine learning for mobile distributed scenarios. This paper proposes a notion of DP for FL for infrastructures with actors and policies formalized in the Isabelle Insider and Infrastructure framework (IIIf). To illustrate this extension of the IIIf by FL and DP on a practical example, we apply the extended framework to a case study from GIS systems for extreme weather warning to control privacy.