Gaussian Process Learning for Distributed Sensor Networks under False Data Injection Attacks

Distributed sensor networks are the backbone of many modern intelligent systems. The collected sensor data are fused and processed to learn the underlying model (such as the Gaussian process) of interested physical process, which serves as critical knowledge for the further decision making process. On the other hand, the distributed and heterogeneous nature of connected devices makes them vulnerable to cybersecurity attacks. One common type of attacks is the false data injection attack, which is implemented by hijacking nodes and modifying sensor measurements. In this work, we study the problem of Gaussian process learning for distributed sensor networks under false data injection attacks. The proposed algorithm is based on formulating consensus for the unknown hyper-parameters of the Gaussian process over the network, where statistical measures of the reliability of the local maximum likelihood estimates are used as the weights in the consensus formulation. Furthermore, we investigate the impact of choosing different subset of nodes to deploy the FDIA, based on the topological properties of the subset of nodes. The simulation result shows that extra effort must be invested to protect the integrity of data from nodes with high centrality, due to their critical positions in the information flow over the network.