Antagonistic control

Abstract

In antagonistic control we find an input sequence that maximizes (or at least makes large) an objective that is minimized in typical control. Applications include designing inputs to attack a control system, worst-case analysis of a control system, and security assessment of a control system. The antagonistic control problem is not convex, and so cannot be efficiently solved. We present here a powerful convex-optimization-based heuristic for antagonistic control, based on the convex-concave procedure, which can be used to find bad, if not the global worst-case, inputs. We also give an S-procedure-based upper bound for antagonistic control, applicable in cases when the objective and constraints can be described by quadratic inequalities, and use this to verify on examples that our method yields inputs very close to the (global) worst-case.

Estimating the Impact of Cyber-Attack Strategies for Stochastic Networked Control Systems

“…The case N z < N r captures attacks that maximize damage in N z steps, and prevent the defender noticing this in additional N r − N z steps. The case N z > N r models ambush attacks [37], where the attacker stealthily prepares N r steps, and then launches a not necessarily stealthy attack in the remaining time. Although we focus on the case N r = N z = N , the analysis that follows can be extended to cover the aforementioned cases as well.…”

Adversarial Model Predictive Control via Second-Order Cone Programming

“…The work was supported by ARO through contract W911NF-17-1-0092, US DoE EERE award de-ee0008006, and NSF through grants CNS 1544771, EPCN 1711188, AMPS 1736448, and CAREER 1752362. sought via convex-concave approximations [6], semidefinite relaxations [3], or general nonlinear programming methods. Alternatively, an attacker may avoid the non-convex problem by selecting a target state (which is different from the system's intended operational state) and minimizing deviations from it [7].…”

“…Examples include adding a rogue car to a vehicle platoon [4] or malicious demand response in power grids [5]. Lastly, instead of injecting false sensor data or introducing adversarial agents, the attacker might take over the whole system and control it with an antagonistic algorithm [6] that maximizes damage.…”

“…This leads to a non-convex problem which is NP-hard in general. Due to the computational complexity, suboptimal solutions are typically sought via convex-concave approximations [6], semidefinite relaxations [3], or general nonlinear programming methods. Alternatively, an attacker may avoid the non-convex problem by selecting a target state (which is different from the system's intended operational state) and minimizing deviations from it [7].…”