Adaptive Client Participation in Budget-Constrained Federated Learning with Extreme Label Skew

Federated Learning (FL) is a distributed machine learning technique in which multiple clients collaboratively train a global classification model while ensuring that private data stays decentralized. However, under skewed label distributions, FL struggles to achieve high accuracy. The problem is further exacerbated when clients have limited energy budgets that prevent their full participation during the training flow. To address these limitations, we propose FL with Adaptive Concurrency via Gradient Feedback (FedAGF), a control strategy that dynamically adapts the number of clients selected for synchronization. FedAGF implements continuous monitoring of the global model updates and increases concurrency when needed, that is, when learning stalls, optimizing the allocation of clients' budgets. Experimental results on the CIFAR-10 and CIFAR-100 datasets demonstrate that FedAGF improves accuracy by up to 14.14% compared to the existing methods.