Process Pattern-Based Flexible Federated Learning

Federated Learning (FL) is an increasingly adopted approach for training Machine Learning models with geographically distributed clients. Multiple clients can benefit from an enlarged cross-location training dataset without sharing any potentially sensitive data, thereby improving data privacy compliance. Various frameworks exist for implementing the basic FL approach. However, the definition and implementation of entire FL systems is cumbersome. In particular, different requirements must be addressed flexibly, as they can differ from use case to use case. Even within a single FL system, requirements may vary between individual clients, for example, in terms of data pre-processing or the execution of training procedures. This paper presents a novel modular approach that incorporates the use of FL frameworks and the specification language Business Process Model and Notation (BPMN). It adopts the concept of process patterns to facilitate the flexible design and implementation of FL systems. For this purpose, the general FL process is modeled in BPMN. The process patterns provide templates for the flexible extension of the general FL process model, for example, the process pattern “Scheduling” for customization of client-specific local training procedures. ... mehrTo further facilitate the implementation of FL systems, this paper demonstrates how the BPMN process is interwoven with an underlying FL framework instance that encapsulates the actual FL. Finally, this paper shows how this approach is applied to two case studies by modeling the FL processes and deploying and executing them in a workflow engine such as Camunda.