Overcoming Data Skew: Adaptive Federated Reparameterization in Edge Intelligence

In next-generation networks, the integration of edge intelligence and large models has transformed real-time applications such as intrusion detection and anomaly monitoring. However, standard federated learning approaches (e.g., FedAvg) struggle with data heterogeneity across edge devices, often yielding suboptimal performance - especially for minority classes. To address this challenge, we propose an adaptive reparameterization framework called FedGRAA (Federated Gradient-based Adaptive Aggregation). FedGRAA employs an adaptive, gradient-driven weighting strategy that dynamically adjusts each edge device's contributions based on its data distribution, ensuring that clients with smaller datasets still wield a meaningful influence on the global model. This design not only enhances detection accuracy for underrepresented classes but also preserves data privacy by keeping raw data local to the devices. Experimental results show that FedGRAA outperforms FedAvg, increasing accuracy from 80% to 88.88% for minority-class detection. Our findings highlight the importance of adaptive reparameterization in advancing federated learning for edge devices within the evolving landscape of next-generation networks.