In modern autonomous systems, from quadrotors to self-driving cars, the controller is only as good as the data it receives. Traditional state estimation, typically handled by Kalman Filters, assumes that sensor noise is stochastic and benign. However, an intelligent adversary can inject "stealthy" false data—signals designed to look plausible while slowly leading the system into a catastrophic state.
The research direction emerging from these papers focuses on Resilient State Estimation: the ability to maintain accurate situational awareness and operational safety even when a subset of sensors or actuators is maliciously compromised.
The Combinatorial Challenge of Trust
The core difficulty in resilient estimation is not just filtering noise, but identifying which sensors have been hijacked. If a system has $N$ sensors and an attacker can compromise up to $K$ of them, there are $\binom{N}{K}$ possible attack patterns.
To solve this, researchers like Yasser Shoukry, Paulo Tabuada, and Sanjit A. Seshia have pioneered the use of Satisfiability Modulo Theory (SMT) solvers. In Secure State Estimation Under Sensor Attacks: A Satisfiability Modulo Theory Approach, they frame the problem as a search for the largest subset of sensors that are mutually consistent with the system's physical dynamics. A SAT solver handles the combinatorial explosion of sensor combinations, while a "theory solver" checks if the chosen subset satisfies the underlying physics of the system.
Redundancy Through Multi-Estimator Banks
When an attack is "stealthy"—meaning it stays within the statistical bounds of normal noise—a single observer might be slowly "pulled" away from the truth. To counter this, Luyao Niu and Andrew Clark propose maintaining a bank of estimators.
Each estimator in the bank processes a different subset of sensors. By solving a quadratically constrained quadratic program (QCQP) across these multiple perspectives, the system can find a control input that is safe for all likely attack scenarios. This ensures that even if the system doesn't know exactly which sensor is lying, it can still reach its goal with provable safety guarantees.
Detection and the "Escape Time" Strategy
For highly dynamic systems like UAVs, detection is only the first step. Once a GPS spoofing attack is identified—perhaps via a transformer-based detector like QUADFormer Jialu Li—the system must immediately transition to a resilient mode.
Wenbin Wan and Petros G. Voulgaris describe a "dual-mode" observer. Upon detection, the UAV switches from external GPS fusion to internal-only "dead-reckoning" using IMUs. While this prevents the attacker from leading the drone astray, internal sensors suffer from drift. This introduces the concept of Escape Time: the finite window during which the UAV can perform an emergency maneuver to exit the attacker's range before the accumulating drift makes navigation impossible.
Bypassing Sensors with Physical Cross-Checks
Recent work has moved toward using non-traditional data to reconstruct state. The MARS framework by Haocheng Meng and colleagues demonstrates that if inertial sensors (IMUs) are attacked, the system can reconstruct its state using rotor speeds and motor dynamics. Similarly, Feilong Wang and Xuegang Ban utilize Roadside Units (RSUs) to provide an external "ground truth" that can re-initialize a vehicle's state, purging the corrupted history from the estimator's memory.
By treating the system's physical laws as the ultimate source of truth, these frameworks ensure that even when the "eyes" of the system (the sensors) are compromised, the "brain" (the estimator) can still deduce the reality of its motion.