Optimization of Graph Neural Networks for Real-Time Intrusion Detection in Dynamic Mobile Ad-Hoc Networks

Abstract

AD-HOC Mobile Networks (Manets) face significant security challenges due to their dynamic topology and vulnerability to sophisticated cyber threats. Traditional IDS detection (IDS) systems usually cease to provide real -time adaptive protection against evolutionary attacks on these resource restriction environments. This article presents an optimized graphic neural network IDS (GNN-IDS IDS) that addresses these limitations through three main innovations: dynamic graphs representation learning, light architecture design and online adaptation mechanisms. Our approach reaches 93.2%detection accuracy-SUPPORT LSTM and signature-based methods at 4.6-20.7%, maintaining real-time performance (28ms latency) suitable for mission critical applications. By incorporating contrastive learning and opponent training, the system demonstrates exceptional robustness, improving zero-day attack detection (F1: 0.83) and reducing evasion success rates to just 12%. Extensive evaluations in simulated test scenes (NS-3/OMNET ++) (Raspberry PI/ESP32) confirm the practicality of the solution, showing scalability to over 500 knots with only 8% energy energy-critical advantage for battery-dependent manet deployments. GNN IDS adapt to topology changes in 2.1 seconds, exceeding static GCNs in 2.7 × in dynamic scenarios. These advances not only improve Manet safety, but also provide a structure to protect other dynamic networks such as 5G/6G and satellite constellations.