FedEPL+: Carbon-Aware Client Selection With Valid Differential Privacy in Federated Learning

Problem

Federated learning must choose clients for training while respecting privacy requirements, device heterogeneity, and the environmental cost of distributed computation.

Why It Matters

Privacy-preserving AI still consumes energy. Sustainable federated learning needs to treat carbon cost as a first-class system constraint, not an afterthought.

Core Idea

FedEPL+ frames client selection as a carbon-aware decision problem while preserving valid differential privacy guarantees during federated training.

Main Result

The work is listed as a 2026 IEEE conference paper on carbon-aware client selection with valid differential privacy in federated learning.

What Was Difficult

The method has to balance model quality, privacy accounting, client availability, carbon signals, and fairness across distributed participants.

Resources

Open Directions

• How should carbon signals be estimated for distributed clients?
• Can privacy and sustainability constraints be optimized jointly?
• Does carbon-aware selection affect fairness across participants?