Robust federated incremental learning in heterogeneous environments

Federated Learning (FL) enables collaborative model training across distributed participants without sharing raw data, thereby preserving privacy and ensuring regulatory compliance. However, the robustness of FL is severely challenged in industrial scenarios, where data distributions are often non-independent and identically distributed (non-IID) and malicious clients may launch Byzantine attacks. Existing Byzantine-robust FL methods typically rely on static defense strategies, such as rule-based aggregation or outlier detection, which are insufficient for dynamic and heterogeneous environments. To overcome these limitations, this paper proposes a framework for robust federated incremental learning in heterogeneous environments (RFIL). The proposed method integrates incremental learning with dynamic client selection to enhance resilience against adversarial behaviors. Specifically, we calculate pairwise euclidean distances among model nodes and apply a pearson correlation coefficient based analysis to characterize client updates. Suspicious or unreliable updates are filtered before aggregation, which improves the robustness of the global model. Extensive experiments conducted under heterogeneous and adversarial settings demonstrate that our approach achieves higher robustness and adaptability compared with existing Byzantine-robust FL frameworks. These results highlight the potential of federated incremental learning as a practical solution for secure and reliable collaborative modeling in real-world industrial environments.