Detecting Poisoning Attacks in UAVs via Spatio-Temporal Graph Neural Network and Walrus Optimizer

Unmanned Aerial Vehicles (UAVs) are now increasingly used in autonomous navigation tasks where accurate decision-making and safe operation are crucial. However, the autonomous decision-making system faces many threats of adversarial attacks, such as data poisoning and Trojan trigger attacks that could compromise its safe operation. In this work, we propose a Spatial-Temporal Fusion Graph Neural Network with a Walrus Optimizer (STFGNNet-WO) to improve UAV navigation with integrated visual data and control signals into its predictions that can help mitigate those attacks. The process begins with collection of high-quality data from two types of datasets. First, steering angle data will be collected from the Udacity dataset, and second, collision probability data will be collected from our custom GoPro dataset. These datasets include substantial and complementary visual inputs and control signals, which are important for UAV navigation tasks. Next, the Spatial-Temporal Fusion Graph Neural Network (STFGNNet) extracts spatial-temporal features by fusing adjacency matrices and gated CNN outputs. Finally, the Privacy Preserving Federated Learning (PPFL) framework is deployed, enabling collaborative model training across UAVs while safeguarding data privacy and filtering out poisoned updates. The model achieved an RMSE of 0.105, 99% sensitivity, specificity, and precision, and 99.2% accuracy, outperforming existing methods in robustness and accuracy.