Attack-tolerant Trajectory Prediction using Generative Adversarial Network Secured by Blockchain Application to the UAS-S4 Ehécatl

Abstract

A robust data-driven algorithm is designed for Aircraft Trajectory Prediction (ATP). A Neural Network model predicts future trajectories of aircraft relying on the input vector containing latitude, longitude, altitude, heading, speed, and time. The model is constructed based on Generative Adversarial Networks (GANs) architecture. The GAN model is highly robust against Adversarial Attacks due to its inherent generative feature. Blockchain is used as a Ledger Technology (LT) in order to trustworthy store the legitimate predicted values that are used for further predictions. In other words, blocks refuse storage of adversarial predicted values as they are detected as adversarial samples and are not approved by the Blockchain. For validation studies, trajectories for training the GAN model were generated using our UAS-S4 Ehécatl simulation model. Adversarial Attack Tolerance based on fooling rates was considered as performance index. The obtained results confirmed the excellent effectiveness of our Blockchain-secured GAN in the case of adversarial white-box and black-box attacks.