FedLucra : Federated Learning With Loss‐Utility Controlled Robust Aggregation

ABSTRACT Federated learning enables collaborative model training without centralizing private data. However, conventional aggregation strategies generally determine client contributions primarily according to local data volume and do not directly consider differences in local model performance. To address this limitation, we propose FedLucra, a modular loss‐aware aggregation framework that jointly uses client validation loss and local data volume to determine aggregation weights. Because loss‐derived weights may vary substantially across clients and communication rounds, FedLucra incorporates Exponential Moving Average smoothing, Interquartile Range‐based score filtering, and an optional FedAvg warm‐up phase to improve aggregation stability. We further introduce Loss‐Weighted Differential Privacy and Adaptive Loss‐Weighted Differential Privacy, which transform client‐specific loss‐utility scores into bounded privacy budgets, with the adaptive variant adjusting its allocation parameters according to the round‐wise distribution of client scores. FedLucra is evaluated on MNIST, Fashion‐MNIST, and CIFAR‐10 under IID and Dirichlet non‐IID distributions with different levels of data heterogeneity and federation size, including partial client participation. Additional experiments are conducted on the real‐world State Grid Corporation of China electricity‐theft dataset, while the proposed privacy mechanisms are evaluated under multiple privacy configurations. The results show that EMA smoothing and IQR filtering significantly strengthen inverse‐loss aggregation and enable FedLucra to maintain competitive and stable performance across diverse experimental conditions. Moreover, the proposed loss‐aware privacy mechanisms can improve model utility relative to uniform‐budget differential privacy in several evaluated settings, with the observed gains depending on the dataset, aggregation method, and allocation configuration. Overall, FedLucra provides an interpretable and flexible framework for combining stable loss‐aware aggregation with client‐specific privacy‐budget allocation in federated learning.