FedDriftGuard adaptive federated learning with differential privacy for concept drift in edge environments

Federated learning (FL) has become a highly promising paradigm for privacy-preserving distributed model training by enabling edge devices to train without sharing raw data. But in practice, edge environments are both non-stationary and asymmetric, with varying data distributions due to shifts in user behaviour, sensing conditions, and overall environmental dynamics. This causes concept drift (sudden, gradual, and recurrent), leading to poor model performance, slower convergence, and predictive bias. Current approaches to FL are not combined to tackle problems of drift adaptation, differential privacy (DP) and resource efficiency (FedAvg, DP-FedAvg). To address these constraints, we present FedDriftGuard. This Federated learning layer unifies client-level drift detection, drift-adaptive aggregation, and adaptable differential privacy into a single, FLE architecture-compatible system. The proposed DP-DriftNet model implements attention-based time encoding to capture changing data patterns and drift-directed feature weighting to allow greater flexibility in the presence of distributional changes. A drift-optimal privacy scheduler allocates noise probabilistically, subject to a limited privacy budget, thereby enforcing an appropriate privacy-utility trade-off without cancelling formal DP guarantees. Also, update sparsification, compression and periodic transmission techniques are used to reduce communication overhead. Decades of experimentation on real-world and synthetic drift datasets have shown that FedDriftGuard outperforms baseline FL techniques, achieving accuracy and F1-score gains of 9-14% and 11-17%, respectively, with adaptation latency 28% shorter and communication cost 20-35% lower. Such findings are statistically significant and confirm the soundness of the suggested method. FedDriftGuard offers effective, scalable privacy-preserving learning in adaptable, edge-drifting environments.