Autonomous Navigation in Dynamic Maze Environments Under Adversarial Sensor Attack

Autonomous navigation of complex and dynamically changing environments is challenging, but when the sensors that determine observables is unreliable, reliability is undermined. Current research largely focus on either efficient navigation with reliable sensors, or sensor fusion mechanisms to offer robust position estimation. In many sensor fusion cases, it assumes that there is a regular structure to the environment (e.g., road or cityscape) or that the attacks are intermittent, and hence leverages on Kalman filters or alternative sensors as a holding fusion mechanism. Here, we present work on reliable UAS navigation under adversarial sensor attack (e.g., persistent wide area or target following attacks) in irregular environments (e.g., a dynamic maze). We cannot rely on regularized routes in training, nor can we rely on a holding fusion algorithm to wait out an attack whilst simply estimating the next position in a smooth trajectory. The autonomous platform in our work needs to make rapid assessments and manoeuvres in the maze to reach its goal. In this work, we present initial findings on the effect of two types of attack on a reinforcement learning (RL) controlled drone: (1) a laser attack on the camera sensor, and (2) an area of effect denial attack on the lidar. We develop Variational autoencoder (VAE) based sensor filtering approaches to detect the effect of attacks with 95% success and assist in achieving robust navigation.