Structural Causal Models for Distributional Shift in Federated Learning

Purpose and goal

This project aims to address differing data distributions and client drift in federated learning by developing solutions that can handle distribution shifts and improve model robustness and generalization. The project will also explore estimating uncertainty in federated learning models and integrating causal inference to mitigate confounding bias. The methods will be tested on real-world telecommunication datasets. The goal is to create more resilient FL systems capable of handling dynamic data sources.

Expected effects and result

This project will create technology that can mitigate confounding bias in federated learning. The expected results is at least two scientific publications in highly rated computer science conferences or journals, and a demonstrator that can be run on relevant datasets. Furthermore, the demonstrator will be tested and evaluated on real world datasets from the telecom industry.

Planned approach and implementation

This project is organized around three work packages which all will run for the duration of the project. WP1: Project administration and dissemination 97h, WP2: Structural causal models for distribution shift in federated learning 920h, and WP3: Pathway to impact 220h. In WP2, we will develop solutions to enhance the robustness and performance of FL. This includes mitigating distribution shift, uncertainty estimation in FL models, and leveraging causal reasoning techniques to reduce the impact of distributional shifts. In WP3, the techniques will be tested on real world problems.