Learning dynamic jamming models in cognitive radios

Abstract

Cognitive radio (CR) integrates results from software-defined radio (SDR), machine learning (ML), and neuroscience for smart radio transmission devices. SDR enables devices to be digitally and dynamically configured in online applications; methodologies and techniques developed to introduce self-awareness in existing systems can be based on ML. Specifically, CR can adaptively regulate its internal parameters in response to the changes in the surrounding environment. New physical layer security issues are also emerging, for example, smart jamming attacks aim to reduce the quality of service or to disrupt legitimate communications. In this context, the electromagnetic spectrum represents the environment, while signals inside it are the individual entities. A CR-to-spectrum interaction consists of a dynamic process that can be driven by a CR device. Learning dynamic and measurement models from spectrum data is the main objective in CR applications. To learn a model, statistical signal processing techniques can be used. Such models can be considered as parametric Bayesian filters that allow a CR to estimate current state of observed entities (including CR itself) and to predict their actions in the near future. Adaptive hierarchical Bayesian filters able to cover nonstationary entity behaviors can be described through probabilistic graphical models (PGM). Interacting entities can be modelled by coupling multiple PGMs related to different entities. In this chapter, state-of-the-art on representation and learning of dynamic models for physical layer security is introduced along with some future directions. An experimental framework is then presented with two currently investigated applications: Spectrum Intelligence and TV White Spaces (TVWS).