Abstract
The 5G massive multiple-input multiple-output (MIMO) approach not only increases spectral efficiency levels efficiency, and yet also enables location-based services (LBS), such as outdoor vehicle localization and inside user localization, with the deployment of 5G communication systems. Currently, massive MIMO localization utilizing channel state information (CSI) or angle-delay profiles which is often called as ADP is done with a deep convolutional neural network (DCNN). However, the DCNN model's robustness for huge MIMO localization is not investigated.
In the project we will understand the effects of adversarial attack and defense (i.e., adversarial training) on massive MIMO localization using the DCNN model and neural ordinary differential equation (ODE) model and look at the performance of these models in a dynamic environment. This research starts with the introduction to massive MIMO system including the channel model and ADP creation. Then, we discuss the DCNN model and the neural ODE model for massive MIMO localization, respectively. In addition, three white-box attacks and adversarial training are discussed. Finally, experimental results show the proposed neural ODE with adversarial training could greatly improve the robustness of massive MIMO localization in indoor, outdoor, and dynamic environments. TensorFlow, a Python framework, was used to implement the project, which utilizes hardware accelerators like GPUs. MATLAB was used to generate the dataset, and Google Colab Pro was used to train the models on this dataset.