PCA-Featured Transformer for Jamming Detection in 5G UAV Networks

Uncrewed Aerial Vehicles (UAVs) face significant security risks from jamming attacks, which can compromise network functionality. Traditional detection methods often fall short against AI-powered jamming that dynamically adapts its behavior, while contemporary Machine Learning (ML) approaches frequently require extensive feature engineering and struggle to capture temporal patterns in attack signatures. The vulnerability extends to 5G networks employing Time Division Duplex (TDD) or Frequency Division Duplex (FDD), where service quality can deteriorate under deliberate interference. We introduce a novel U-shaped transformer architecture that leverages Principal Component Analysis (PCA) to refine feature representations for improved wireless security. The training process is regularized by incorporating output entropy into the loss function, inspired by the Soft Actor-Critic (SAC) algorithm in Reinforcement Learning (RL), to enhance robustness against jamming attacks. The architecture employs a modified transformer encoder tailored for wireless signal features, including Received Signal Strength Indicator (RSSI) and Signal-to-Interference-plus-Noise Ratio (SINR), together with a custom positional encoding mechanism that captures the inherent periodicity of wireless signals for improved temporal modeling. To further optimize training, we propose a batch size scheduler and apply chunking techniques to accelerate convergence on time-series data, achieving up to a tenfold improvement in training speed within our U-shaped encoder–decoder transformer. Experimental evaluations demonstrate the effectiveness of the proposed entropy-based framework, achieving detection rates of 89.46% in Line-of-Sight (LoS) and 85.06% in Non-Line-of Sight (NLoS), with F1-scores up to 0.90 and Area under the Curve (AUC) values of 0.97 and 0.92 respectively, underscoring the model’s balanced performance across key evaluation metrics. The method significantly outperforms existing solutions, surpassing XGBoost (XGB) by approximately 4.5% and other Deep Learning (DL) approaches by more than 2%.