Abstract
The increasing reliance on drone swarms for various applications necessitates robust real time anomaly detection mechanisms to ensure operational security and efficiency. Federated learning is particularly well-suited in the drone context as it enables decentralized data processing, preserving data privacy and security while enhancing detection accuracy. In this paper, we explore optimization methods for federated learning-enabled network incident anomaly detection in drone swarms using the NSF AERPAW platform. To achieve this, we demonstrate a defense mechanisms such as differential privacy and adversarial training, to strengthen the robustness of federated learning models against data poisoning attacks. We are collecting three sets of metrics for accuracy, system and network usage under a variety of test cases reflecting benign, mildly poisoned, highly malicious, safe situations. The experimental results reveal that adversarial training is particularly effective, achieving up to 91.1% accuracy with a 33% data poisoning volume. Additionally, we evaluate the computational overhead introduced by these defenses, finding that while they enhance security, they also increase CPU usage by up to 233% in active drone scenarios. These findings highlight the trade-offs between security and operational efficiency in FL-enabled drone swarms, offering critical insights for deploying robust, real-time anomaly detection systems in decentralized environments.