South Dakota State University
Securing Smart Agriculture with Communication-efficient Federated Unlearning
Abstract
dc:description.abstractThis thesis addresses the data poisoning attack in smart agricultural systems using federated unlearning, where the contribution of compromised field devices must be removed from the global model under communication constraints. We propose FedSCAN, a three-phase framework that identifies critical layers based on parameter sensitivity, classifies active neurons within these layers through relative weight change analysis, and performs unlearning via sparse low-rank adaptation to active neurons while freezing base model parameters. FedSCAN limits computational overhead to a compact parameter subset, enabling devices to participate in a communication-efficient manner. Experimental results on four datasets demonstrate that FedSCAN achieves up to 426.3× communication cost reduction compared to retraining while maintaining remaining accuracy within 3% of the retraining baseline. Compared with state-of-the-art methods, FedAU, FedEraser, and FedOSD, on the agricultural datasets, FedSCAN reduces communication overhead up to 196.3× while achieving up to 16.4% higher remaining accuracy than FedAU and 13.1% higher than FedEraser. On the Fruit-360 dataset with 131 fine-grained categories, FedSCAN achieves up to 181.8× communication reduction while maintaining remaining accuracy within 0.07% of the retraining baseline.
Degree
thesis:*- Name thesis:degree_name
- Master of Science (MS)
- Level thesis:degree_level
- Thesis - Open Access
- Discipline thesis:degree_discipline
- Electrical Engineering and Computer Science
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Pudasaini, Ujjwal
- Contributors dc:contributor
-
- Jun Huang
Rights
dc:rightsIdentifiers
dc:identifier.*- Repository record dc:identifier
- https://openprairie.sdstate.edu/etd2/2135