MoARNN-AM: Multi-Objective Automated Recurrent Neural Network With Attention Mechanism for Cyber-Attack Detection of UAV

As one of consumer unmanned electronic systems, unmanned aerial vehicles (UAVs) have become ubiquitous in daily life, essential for numerous tasks, and will play a pivotal role in future wireless networks and Internet-of-Things. However, their widespread adoption and connectivity make them vulnerable to various cyber threats. Although deep learning-based attack detection models offer promising solutions for enhancing UAV network security, these models typically rely on manual trial-and-error approaches for determining hyper-parameters and neural architectures, resulting in limited generalization capability and often overlooking model lightweightness. To address these limitations, this paper proposes an innovative automated multi-objective recurrent neural network (RNN) with attention mechanism, called MoARNN-AM, to effectively solve attack detection problems in UAV systems. In MoARNN-AM, we consider six typical RNN variants and three widely-used attention mechanisms as core classification models for feature extraction and data learning of UAV systems. First, an effective encoding mechanism is developed to represent different combinations along with their corresponding hyper-parameters and neural architectures. Subsequently, considering both attack detection performance and model lightweightness as two objectives, we elaborately design an efficient non-dominated sorting genetic algorithm II (NSGA-II)- based evolutionary operation to evolve various combinations with their associated hyper-parameters and neural architectures for discovering optimized RNN with attention mechanism model. The performance of the proposed MoARNN-AM method is validated using two datasets, i.e., UAV-INDD dataset and WSN-DS dataset collected from different UAV systems. Experimental results demonstrate that MoARNN-AM outperforms five state-of-the-art manually designed attack detection models in terms of accuracy, precision, recall, and F1-score metrics while maintaining superior model lightweightness.