Detecting integrity attacks on control systems using robust physical watermarking

Abstract

Ensuring the security of control systems against integrity attacks is a major challenge. Due to the events of Stuxnet, replay attacks in particular have been considered by the research community. Replaying previous measurements of a system in steady state allows an adversary to generate statistically correct virtual outputs which can bypass traditional detectors. The adversary can then inject destabilizing inputs to cause damage to the plant. The method of injecting secret noisy control inputs, or physical watermarking, has recently been proposed to detect replay attacks. However, the proposed watermarking design methods assume that the adversary does not use his potential access to real time communication channels to create stealthy virtual outputs to send to the defender. In this paper, we formulate an attack model for an adversary who uses knowledge of the system as well as access to a subset of real time control inputs and sensor outputs to construct stealthy virtual outputs. A robust physical watermark and detector to counter such an adversary is proposed. 53rd IEEE Conference on Decision and Control

Dynamic watermarking for general LTI systems

“…However, CPS frequently has a decentralized structure; and so approaches that detect attacks by decoupling different sensing and actuating components of the system are particularly useful for ensuring safe operation. This paper designs a dynamic watermarking approach for detecting malicious sensor attacks for general LTI systems, and has two main contributions: First, we generalize the watermarking approach developed in [8] with partial state observations and MIMO LTI systems with a full rank input matrix and full state observations under an arbitrary attack, and our generalization applies to general LTI systems under a specific attack model that is more general than replay attacks [9]. Second, we show that modeling is important for designing watermarking techniques: For instance, dynamic watermarking was used to detect sensor attacks in an intelligent transportation system [10]; however, here we show that persistent disturbances such as those from wind can invalidate watermarking approaches, and we propose an approach based on the internal model principle to compensate for persistent disturbances.…”

Statistical Watermarking for Networked Control Systems

“…Next observe the upper block triangle of (19) is correct by construction of Q I,J,k i,j and by definition of the e i , and so we only have to prove that the lower-right block is correct. In particular, note that E((C j δ i,n − z n )(C j δ i,n − z n ) T ) = E((C j δ i,n )(C j δ i,n ) T )+Σ Z,I since z i,n is independent of δ i,n by (11). This implies that we have that as-lim n E((C j δ i,n − z n )(C j δ i,n − z n ) T ) = C j D i C T j + Σ Z,I .…”

Simulation and Real-World Evaluation of Attack Detection Schemes

“…Initially, Dynamic Watermarking was developed for LTI systems with full rank input matrices and full state observations. These methods were proven to detect arbitrary attack models [15] including attacks that replayed a measured signal [16]. These methods were later extended to generalized LTI systems [17], and to networked control systems [18].…”