InteFL Framework: Optimizing Federated Learning with Metacognition for Application Design and Deployment

We argue that metacognition should be the guiding principle for designing and deployment of Federated Learning (FL) applications. The metacognitive features in FL, such as monitoring and control, allow for the dynamic adaptation to the environment, client updates, enhanced user interaction, and optimized resource-aware learning. We introduce InteFL, the first framework that facilitates designing metacognitive FL systems by providing mechanisms and tools to investigate how possible changes in execution and learning conditions impact FL performance and security, and then to choose or optimize the learning structure and process based on their evaluation. We demonstrate the framework application in practice on the cases showing how the choice of FL hyperparameters such as model aggregation algorithms can be dynamically adapted or fine-tuned in response to changing execution environment, including data quality variations, network conditions, or adversarial attacks.