A Novel Technique to Handle Concept Drift with Dynamic Minority Oversampling and Adaptive Client Weighting in Federated Learning

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

In a federated environment, class imbalance and concept drift often arise because privacy concerns prevent sharing raw data. No existing state-of-the-art method resolves these three challenges simultaneously. We propose the Drift-aware Federated Learning (DAFL) framework, which performs the KS test, applies dynamic SMOTE, and uses adaptive weighting to address all three challenges simultaneously. Experimental results demonstrate that on the Credit Card Fraud dataset with extreme class imbalance (IR=578), all baseline methods achieve a low G-Mean (below 0.43). In contrast, DAFL achieves a G-Mean of 0.5774, confirming its superiority in imbalance scenarios. We also ran additional tests on our model on the Credit card fraud and UNSW datasets for ablation and Wilcoxon tests to further demonstrate the statistical validity of the framework.