AutoML-DAFL : A Drift-Aware Federated Learning Framework for Cyber-Physical Aquaculture Water Quality Monitoring

Abstract

Federated Learning (FL) for Internet of Things (IoT) and cyber-physical sensor networks, such as aquaculture water monitoring, faces critical challenges due to temporal sensor drift, non-IID data distributions, and communication constraints across edge devices, which compromise global model stability and resource efficiency. However, most existing federated approaches lack coherent mechanisms to address these issues, leading to degraded performance in realistic edge deployments. We present AutoML-DAFL, a drift-aware, AutoML-guided federated learning framework with a multi-objective reward controller that jointly optimizes predictive accuracy, model consistency, and communication efficiency. The framework integrates temporal drift detection and mitigation into the federated training loop through MAE-based regularization while ensuring persistent convergence. To evaluate its effectiveness, we benchmark AutoML-DAFL against FedAvg and advanced baselines, including FedNova and SCAFFOLD. Extensive ablation studies and comparative analysis on real aquaculture monitoring data demonstrate the contribution of each reward component: removing the communication-aware term degrades model consistency, while excluding MAE-based smoothing reduces training stability. The full AutoML-DAFL configuration, integrating all reward components through multi-objective optimization, achieves the lowest RMSE (0.0721), highest R2, and improved fairness across clients, demonstrating strong resilience to non-IID drift and bandwidth constraints. These results highlight the effectiveness of drift-aware AutoML optimization for resource-efficient, stable federated forecasting in cyber-physical monitoring systems.