Learning-based Detection of Stealthy False Data Injection Attack Applied to Cooperative Localization Problem
Eshaan M. Khanapuri, Rajnikant Sharma, Kevin Brink · AIAA SCITECH 2022 Forum · 2022
View Video Presentation: https://doi.org/10.2514/6.2022-2543.vid Security of autonomous, connected UAVs plays an essential role in their far-reaching capability and acceptance. A single adversarial UAV injected with false bearing measurements in the cooperative localization setting can disrupt other UAVs sharing bearing information with the adversarial UAV. This paper focuses on detecting the stealthy false data injection (not detected by statistical tests) attacks on UAVs in the cooperative localization environment. Our approach is decentralized and requires only local bearing measurements of the UAVs. To detect the attacks, we have explored both machine learning and deep learning methods, such as Dynamic Time Wrapping-Nearest Neighbor (DTW-NN) classifier, Fully Connected Deep Neural Network (FCDNN), and Convolutional Neural Network (CNN). We have shown that pre-processing techniques like Gramian Angular Fields (GAF) and Markov Transition fields improve the performance of deep CNN in detecting attacks. Also, three decentralized detection strategies, Unique Neural Network (UNN), General Neural Network (GNN), and One Shot learning, are studied. The results show that we can achieve up to 98$\%$ accuracy with the One-Shot learning approach. The problem can be generalized to $n$ number of UAVs in the environment with various sensing topologies. Finally, the results also show the performance of the strategies and algorithms when the attacker has complete and incomplete information about the system.