Time-Series Modeling-Based Early Detection of DDoS Attacks in Drone Networks

초록

Drone (UAV)-based ad hoc networks are highly vulnerable to Distributed Denial of Service (DDoS) attacks due to their resource constraints and dynamic connectivity. To ensure the survivability of UAVs, ultra-low latency early threat detection is essential. This study proposes three novel time-series network metrics—Packet Flood Rate (PFR), Link Jitter Index (LJI), and Network Congestion Factor (NCF)—optimized for capturing the dynamic characteristics of DDoS attacks in drone networks. To evaluate the effectiveness of the proposed metrics, we applied lightweight deep learning architectures, including 1D-CNN, GRU, and LSTM. The experimental results demonstrate that the 1D-CNN model, guided by the proposed metrics, achieved the highest accuracy with an F1 Score of 0.9669 and an ROC-AUC of 0.9971. Notably, in terms of Average Detection Delay, a critical factor for early defense, the metric-driven 1D-CNN recorded 0.364 steps, reducing the detection time by approximately 30% compared to GRU (0.527) and LSTM (0.522) with statistical significance (p < 0.001, d = 0.6). Furthermore, despite requiring significantly fewer parameters (20,097), the 1D-CNN achieved a per-window inference latency of 0.611 ms on a standard CPU, demonstrating computational efficiency suitable for edge deployment in resource-constrained UAV environments. These results quantitatively demonstrate that the proposed feature-engineering approach combined with lightweight deep learning is highly viable for real-time threat mitigation in resource-constrained UAV networks. © 2026 by the authors.

키워드