Federated Unlearning System Removes Drone Data Without Full Network Retraining

A novel federated unlearning framework, FedUnlearn-HetFL, successfully erased target drone data during collaborative intrusion detection while cutting communication rounds by 95.0 percent and achieving a 1.70-fold execution speedup compared to retraining from scratch. Evaluated on the ISOT Drone benchmark under Dirichlet non-IID partitioning, the framework collapsed target node accuracy from 99.16 percent to 15.98 percent, with the receiver operating characteristic area under the curve dropping from 0.9999 to 0.4842. Concurrently, retained network clients maintained anomaly detection accuracy above 99.2 percent, and black-box membership inference audits confirmed empirical data erasure.

Collaborative intrusion detection across Flying Ad-Hoc Networks allows Unmanned Aerial Vehicles to identify distributed network anomalies, but deploying such systems under privacy regulations like the General Data Protection Regulation Right to be Forgotten faces two primary bottlenecks. First, structural feature heterogeneity stemming from diverse onboard sensor suites causes standard federated learning algorithms like FedAvg to fail. Second, retraining models from scratch imposes a severe computational burden when an individual drone departs the network or requests data retraction.

To handle asymmetric sensor inputs, the framework integrates localized attention-gated autoencoders that project asymmetric client feature dimensions of 40, 45, 50, and 62 features into a unified 64-dimensional latent space coupled with a shared 1D-CNN backbone. When an unlearning request is initiated, FedUnlearn-HetFL executes targeted unlearning via selective layer-frozen gradient ascent on the classification head, resets local input projections, and applies rapid stabilization fine-tuning across retained nodes to support bandwidth- and energy-constrained drone swarms.

Subject of Research: Federated unlearning for feature-heterogeneous drone network anomaly detection

Article Title: A novel federated unlearning framework with feature-heterogeneity-aware aggregation for privacy-preserving UAV anomaly detection

Article References: Ishak, M. K., Ghosh, K. P., Bithi, M., Silamane, J. B., & Hossain, M. A. (2026). A novel federated unlearning framework with feature-heterogeneity-aware aggregation for privacy-preserving UAV anomaly detection. Applied Intelligence, 56(16), Article 546. https://doi.org/10.1007/s10489-026-07570-x

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07570-x

Keywords: unmanned aerial vehicles, flying ad-hoc networks, federated unlearning, anomaly detection, feature heterogeneity, machine learning privacy, gradient ascent, ISOT drone benchmark, novel, federated, unlearning, framework