Abstract Book of the 3rd World Conference on Business, Management, and Economics
Year: 2025
[PDF]
Mitigating Federated Learning Bias in Drone-based Systems
Prof. Dr. Selwyn Piramuthu
ABSTRACT:
With the increase in the adoption of drones for non-military purposes, it is necessary to study the security and privacy aspects of data generated through their on-board sensors. When multiple drones are used, given the sensitive nature of data gathered in each of these drones, federated learning (FL) has been suggested as a possible means to learn concepts for actionable intelligence in such systems. Since only the local models (and not local data) are shared with the central server in FL, sensitive data are somewhat protected from adversaries. As each drone gathers its own sensor-based data, there is a possibility of bias being introduced through data from some of these drones. Strategies for learning from biased data used in centralized learning are not directly applicable in FL with private local data. Previous approaches have used algorithmic debiasing, cryptographic protocols, robust aggregation, and mitigation of bias at client level before aggregation (i.e., locally fair training). While each of these are effective under certain circumstances, they are not effective overall. We attempt to address this through appropriate client as well as attribute selection at global and local levels during each FL learning round.
Keywords: attribute selection, bias, client selection, drone, federated learning