Attack-resilient state estimation in the presence of noise

Abstract

Abstract-We consider the problem of attack-resilient state estimation in the presence of noise. We focus on the most general model for sensor attacks where any signal can be injected via the compromised sensors. An l0-based state estimator that can be formulated as a mixed-integer linear program and its convex relaxation based on the l1 norm are presented. For both l0 and l1-based state estimators, we derive rigorous analytic bounds on the state-estimation errors. We show that the worst-case error is linear with the size of the noise, meaning that the attacker cannot exploit noise and modeling errors to introduce unbounded state-estimation errors. Finally, we show how the presented attack-resilient state estimators can be used for sound attack detection and identification, and provide conditions on the size of attack vectors that will ensure correct identification of compromised sensors.

Secure estimation for cyber‐physical systems with adversarial attacks and unknown inputs: An L2‐gain method

“…Thus, the robustness of the aforementioned decoders/estimators with respect to the unknown inputs becomes a critical issue. Motivated by the pioneering works of Fawzi et al, l 0 ‐ and l 1 ‐based estimators were proposed in the works of Pajic et al for the CPSs with sensor attacks and noises. In the works of Han et al, a general class of convex optimization‐based estimators, as well as the attack‐resilience analysis, was given for the case of static error correction in the presence of noises.…”

“…One should note that a common characteristic of the existing results is that the state estimation error possesses L ∞ – L ∞ (ie, peak‐to‐peak) performance with respect to the noises or disturbances, that is, the bound of the state estimation error is proportional to the magnitude of the noises or disturbances. However, the performance cannot be optimized within the existing framework, which may potentially result in a poor estimation performance.…”

“…Meanwhile, the mismatched part will be suppressed in terms of L 2 ‐gain rejection property. Different from the results where L ∞ – L ∞ performance is adopted and cannot be optimized, the proposed L 2 ‐gain rejection performance index can be minimized as much as possible with the help of the LMI technique. For most of the existing linear and nonlinear unknown input observers, a necessary condition for the construction of such observers is that the observer matching condition is satisfied . However, this condition is hard to satisfy for many practical systems .…”

Fundamental Challenges Toward Making the IoT a Reachable Reality

“…However, the IoT is already unveiling its transformative role in rebooting the fundamental ways of our interaction with the physical world [Stankovic 2014;Pajic et al 2015;Fawzi et al 2014;Clark et al 2016]. For instance, smart grids that consist of numerous networked power plants are already extensively deployed to dynamically capture the spatiotemporal variations in user demands and optimize the power plant management for maximization of the efficiency and robustness; wearable physiological sensors and actuators are actively adopted to monitor the crucial biomarkers and provide prompt medical response to alleviate or reverse potential health risks [Bogdan 2015;Bogdan and Xue 2015;Ghorbani and Bogdan 2013;Xue et al 2016b;Kumar et al 2013]; smart portable devices are able to run highly varied applications with full awareness of user contexts, which is enabled by vastly varied embedded ambience sensors (e.g., gyroscope, accelerometer, GPS, temperature, and lighting sensors) as well as advanced data processing technologies (e.g., data mining, deep learning).…”

Design methodologies for securing cyber-physical systems

“…Hence, it is necessary to provide nonoptimal methods for attack-resilient control with formal resilience guarantees. For example, in [36] we show how to exploit techniques from compressed sensing to investigate conditions that will enable the use of convex estimators for attack-identification while providing formal resiliency guarantees. Note that since extracting accurate-enough dynamical models for some CPS (e.g., patient modeling in Medical CPS) is quite challenging (if at all possible), there are limitations to the use of model-driven methods for attack-resilient control.…”