Quickest detection of false data injection attack in distributed process tracking

This paper addresses the problem of detecting false data injection (FDI) attacks in distributed sensor networks without a fusion center, where agent nodes are interconnected via a communication graph and each employs a Kalman consensus information filter (KCIF) to estimate a global process. The state estimate at any sensor is computed from both the local observation history and information exchanged with its neighbors, enabling distributed information fusion. An adversary corrupts the observation at an unknown node and time. We propose quickest change detection (QCD) algorithms for both Bayesian (with known prior on attack time) and non-Bayesian (arbitrary attack time) settings, tailored to the non-i.i.d. nature of distributed consensus estimates. Importantly, the detection strategy relies on consensus estimates rather than the innovation, marking a departure from conventional approaches. In the Bayesian case, we develop a recursive computation of the non-trivial detection statistic at each node, despite the non-i.i.d. observations. For the non-Bayesian scenario, we employ a multiple hypothesis sequential probability ratio test for detection and identification, along with a window-limited generalized likelihood ratio (WL-GLR) algorithm for unknown attack strategies. Numerical results demonstrate that our methods significantly reduce detection delays compared to conventional ${\chi }^2$ detectors, offering improved resilience for distributed tracking systems.