Copula-Driven Learning Techniques for Physical Layer Authentication Using Multimodal Data

In this paper, we present a study on copula-driven learning techniques for physical layer authentication (PLA) in wireless communication, using data from multiple modalities. The collective multimodal data is considered as an attribute vector, which is used as a test statistic for the underlying multi-level hypothesis testing problem of PLA. We consider regular-vine copula-based information fusion approaches across various combinations of these attributes resulting in different architectures, which are tested across different datasets. In particular, we consider three datasets which include a Monte Carlo simulations-based synthetic dataset, and publicly available automotive factory (AF) and open area test site (OATS) datasets from National institute of standards and technology (NIST). A comparative performance study of the proposed architectures across the datasets is carried out in terms of detection accuracy. For the classification task, we consider some of the well-known learning algorithms including long short-term memory (LSTM), random forest, K-nearest neighbor, support vector machine and bagging tree techniques. Moreover, we study the effect of correlation introduced across the attributes, and compare it with the case with no correlation among the attributes. Our extensive results provide intra-algorithm, intra-architecture and inter-architecture insights, and show that LSTM offers the best performance, across the datasets and proposed architectures.