FedCogniBand: cognition-aware predictive bandwidth allocation for communication-limited federated learning

Abstract

Federated Learning (FL) facilitates collaborative model training among dispersed clients while safeguarding data privacy; yet, its practical implementation is significantly hindered by substantial uplink bandwidth restrictions and notable data heterogeneity. Even though there have been recent improvements in FL methods that save communication resources, like adaptive sparsity and importance-aware bandwidth allocation, most existing systems still mainly respond to immediate client These methods, which depend exclusively on immediate client metrics like current loss values or gradient norms, neglect to account for temporal client behavior, frequently leading to inefficient bandwidth usage, unstable training dynamics, and postponed enhancements in the performance of the worst client. This study presents FedCogniBand, a smart FL system that predicts and manages communication resources by analyzing and forecasting client risks over time. Each client is defined by a streamlined cognitive profile that encapsulates historical loss dynamics, short-term performance trends, and update stability. The server uses a simple predictive tool to evaluate how clients will perform in the future based on their past cognitive data and manages bandwidth for a short time frame while sticking to a strict overall budget. To improve communication efficiency, FedCogniBand employs semantic-aware sparsity, prioritizing model parameters based on layer-level sensitivity to the worst-client performance instead of raw gradient magnitude. A control method that takes uncertainty into account dynamically balances predictive and reactive allocation. It shows that performance is stable even in noisy or non-stationary environments. Extensive tests on image and network-traffic classification tasks with strict bandwidth limits show that FedCogniBand consistently improves the accuracy of the worst-performing clients, makes convergence more stable, and reduces unnecessary communication compared to top adaptive FL methods. These results show that using a smart, prediction-based way to manage resources creates a reliable and expandable system for FL in situations with very limited resources.