LSTM-Based Early Jamming Threat Detection Scheme for Drone Ad-Hoc Networks

Drone ad-hoc networks are inherently vulnerable to performance-degradation attacks such as jamming, packet disruption, and routing interference due to dynamic topology changes and unstable wireless channels. In such environments, conventional threshold-based detection schemes often fail to identify threats in their early stages because individual performance metrics remain within normal ranges despite emerging abnormal temporal patterns. To address this limitation, this study proposes an LSTM-based early threat detection method that learns the temporal dynamics of network performance indicators, including packet delivery ratio (PDR), connection reliability (CR), and delay. By modeling inter-metric correlations and evolving degradation trends, the proposed approach enables probabilistic inference of abnormal state transitions prior to explicit threshold violations. The proposed method is validated through simulation experiments conducted in a drone ad-hoc network environment under jamming attack scenarios, and its performance is compared with that of conventional threshold-based schemes. The results show that while the threshold-based approach first detected the attack at t = 65 s when predefined metric boundaries were exceeded, the proposed LSTM-based detector identified the attack at t = 45 s with an estimated attack probability of 0.63, achieving approximately 20 s earlier detection. This improvement is attributed to the LSTM’s capability to capture subtle temporal dependencies, directional trends, and cross-metric interactions that precede abrupt metric degradation. Furthermore, the LSTM output probabilities exhibited monotonic growth during the attack period and gradual decay during recovery, indicating robust tracking of network state transitions rather than isolated event detection. These results demonstrate that the proposed method not only enhances early threat awareness but also contributes to resilience-oriented operation by enabling proactive mitigation in drone ad-hoc networks. This study provides quantitative evidence that sequence learning over performance metrics can overcome the structural limitations of threshold-based detection and enable effective early threat detection in drone ad-hoc network environments.