Generalizability of failure impact modifiers

Background: The generalization of findings regarding failure-impact modifiers (FIMs) in Federated Learning is limited by the specific datasets and model architectures used in experimental studies.

Question / Future Work: Validate the generalizability of the derived failure-impact modifiers (FIMs) and their established relationships by extending the experimental analysis to a wider variety of datasets, model architectures, and real-world data properties not covered by the current image, tabular, and time-series examples.

Why It Matters: Establishing the generalizability of FIM interactions is necessary to create broadly applicable fault-tolerance guidelines for cross-silo FL.

Evidence: Transferring the results directly onto other cases is difficult, as they may contain unique properties that were not represented in these cases.