Structural Causal Models for Distributional Shift in Federated Learning

Purpose and goal:

Goals were fulfilled and the project results include new insights on mitigation strategies for FL under distribution shift. Furthermore, the collaboration between RISE and Ericsson has been close and fruitful, and experiments have been carried out on telecom data to study the distribution shifts and the mitigation strategies. One research paper has been submitted for review in a respected journal: On the effects of similarity metrics in decentralized deep learning under distributional shift (Edvin Listo Zec, Tom Hagander, Eric Ihre-Thomason, Sarunas Girdzijauskas).

Expected results and effects:

We have developed experimental strategies to study FL under covariate shifts, both using synthetic data and real-world telecom data. We have developed mitigation strategies that help developers overcome some of the challenges. One research paper has been submitted for review in a respected journal: On the effects of similarity metrics in decentralized deep learning under distributional shift (Edvin Listo Zec, Tom Hagander, Eric Ihre-Thomason, Sarunas Girdzijauskas).

Approach and implementation:

The project has been carried out in close collaboration between RISE and Ericsson. We have had monthly standing meetings and some extra workshops where data and applications have been discussed. RISE has worked on the mitigation strategies and Ericsson has explored them in real world settings.