ATD: Algorithms for Anomaly Detection Using Graphical Models

点击翻译按钮获取中文摘要

The increased connectivity of our physical and virtual environments results in elevated risks from agents who utilize interconnectedness in order to spread and communicate. The PI will develop a mathematical methodology and algorithms to detect corrupt agents / elements in interconnected and dynamically changing environments. Potential applications include the detection of corrupt agents in governmental and financial networks as well as detection of anomalies in electrical networks, chemical plants and other physical infrastructure and detection of election tampering. The common theme of the project is studying systems whose states are defined by graphs and where the detection of an anomaly is based on graph properties and graph algorithms. The PI will study corruption detection on a graph where nodes represent entities that can examine the status of their neighbors. Which graphs and which algorithms are efficient at recovering corrupt nodes if such exist? In a different direction, the PI will utilize Markov Random Fields as models representing the interaction between individuals with the goal of detecting anomalies from a dynamic perspective. The PI will further study new notions of anomalies in voting, where the PI will utilize the mathematical tools from the theories of noise stability and isoperimetry to define and detect unlikely voting configurations.