THESIS
2018
xviii, 155, that is, xix, 155 pages : illustrations ; 30 cm
Abstract
Private and secure remote state estimation in the context of cyber-physical
systems (CPSs) is studied in this thesis. Monitoring a physical process, a sensor
will forward local state estimates as data packets to a remote estimator over
a vulnerable network, which may be attacked by an intelligent adversary (i.e.,
an eavesdropper or denial-of-service (DoS) attacker discussed in this thesis).
Considering an attacker with different information sets and different destruction
abilities, we leverage Markov decision process (MDP) and stochastic game model
( or competitive MDP) to develop a systematic quantitative decision framework
to protect remote state estimation.
First, regarding CPS privacy, we study the novel active eavesdropping attack
and the optimal attack scheme for an atta...[
Read more ]
Private and secure remote state estimation in the context of cyber-physical
systems (CPSs) is studied in this thesis. Monitoring a physical process, a sensor
will forward local state estimates as data packets to a remote estimator over
a vulnerable network, which may be attacked by an intelligent adversary (i.e.,
an eavesdropper or denial-of-service (DoS) attacker discussed in this thesis).
Considering an attacker with different information sets and different destruction
abilities, we leverage Markov decision process (MDP) and stochastic game model
( or competitive MDP) to develop a systematic quantitative decision framework
to protect remote state estimation.
First, regarding CPS privacy, we study the novel active eavesdropping attack
and the optimal attack scheme for an attacker with a priori knowledge of the
sensor's transmission pattern. Aiming at improving the eavesdropping performance
efficiently, the adversary may adaptively alternate between an eavesdropping
and an active mode. In contrast to eavesdropping, active eavesdropping not only enables the adversary to block the data transfer to the estimator, but
also improves the data reception at the same time. However, launching active
attacks may increase the risk of being detected. As a result, a tradeoff between
eavesdropping performance and stealthiness arises, which is formulated
as a constrained MDP. After deriving a sufficient feasibility condition, we develop
an optimal attack policy for the eavesdropper, the structure of which
is threshold-like, and an algorithm based on a Lagrangian learning method is
proposed to find it.
Second, we address the remote state estimation under jamming threats. Assuming
a posteriori knowledge of the transmission pattern, the attacker will
adopt tactical jamming strategies to degrade the remote estimation accuracy,
which may lead the sensor to adjust its transmission power adaptively. This
interactive process between the sensor and the attacker is studied in the framework
of a stochastic game. Two subclasses of the sensor-attacker game are
investigated. On one hand, the intelligent attacker is capable of conducting its
jamming strategies based on the online information it collects. In this case, we
first discuss the existence of stationary Nash equilibrium (SNE) in order to derive
the optimal defensive/offensive power scheme for the sensor/attacker. We
then present the monotone structure of the optimal strategies. On the other
hand, when the online information is hidden (or incomplete) from the attacker,
the dynamic nature of the history structure introduces additional difficulties in
solving the original problem. Thus, to derive stationary optimal power schemes
for each agent, we convert the original game into a continuous-state stochastic
game and discuss the existence of optimal transmission/jamming power strategies.
Finally, we note the importance of online information in achieving CPS security,
and propose a potential countermeasure to DoS attacks: the sensor may
adopt a deception-based (that is, modify the feedback acknowledgements intentionally
to confuse the attacker) transmission strategy. We begin with an assumption that the attacker follows a pre-determined interference scheme without
knowledge of such "tricks". It is revealed by MDP and majorization theory
that the optimal deception scheme for the sensor is consistently cheating, which
is intuitive and easy to implement. However, the situation is more complicated
when a tactical attacker is involved, which may discover this consistent
cheating and modify its attack scheme. To cope with this, we model the strategic
interaction between the sensor and the attacker by a stochastic game with
an asymmetric information structure, and analyze the existence of its optimal
solution after converting it into a belief-based dynamic game.
Post a Comment