176 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO.

1, JANUARY 2021

Secure State Estimation and Control of

Cyber-Physical Systems: A Survey
Derui Ding , Senior Member, IEEE, Qing-Long Han , Fellow, IEEE, Xiaohua Ge , Member, IEEE,
and Jun Wang , Fellow, IEEE

Abstract—Cyber-physical systems (CPSs) empower the inte- theoretical investigations and engineering applications. To
gration of physical processes and cyber infrastructure with the enable various monitoring and control operations, CPSs
aid of ubiquitous computation resources and communication employ networks of multifunctional sensors and actuators as
capabilities. CPSs have permeated modern society and found
extensive applications in a wide variety of areas, including energy, well as cyber components, such as computing and commu-
transportation, advanced manufacturing, and medical health. nication devices, which are closely linked together via some
The security of CPSs against cyberattacks has been regarded shared communication networks, either wired or wireless, to
as a long-standing concern. However, CPSs suffer from extend- accomplish data sensing, collecting, processing, and transmit-
able vulnerabilities that are beyond classical networked systems ting tasks. CPSs are regarded as the heart of the modern
due to the tight integration of cyber and physical components.
Sophisticated and malicious cyberattacks continue to emerge to industry and economic society, and find wide and exten-
adversely impact CPS operation, resulting in performance degra- sive applications, including intelligent transportation systems,
dation, service interruption, and system failure. Secure state automobiles, power/gas/water generation, transmission and
estimation and control technologies play a vital role in warrant- distribution systems, advanced manufacturing systems, and
ing reliable monitoring and operation of safety-critical CPSs. smart cities. CPSs are playing a critical role in the 4th indus-
This article provides a review of the state-of-the-art results for
secure state estimation and control of CPSs. Specifically, the trial revolution, as testified by the Industry 4.0 in Germany
latest development of secure state estimation is summarized in and Industrial Internet in the U.S.
light of different performance indicators and defense strategies. CPSs are essentially complex, large-scale, geographically
Then, the recent results on secure control are discussed and dispersed, and safety-critical. Ubiquitous computing and com-
classified into three categories: 1) centralized secure control; munication resources enable pervasive multilayered CPSs,
2) distributed secure control; and 3) resource-aware secure con-
trol. Furthermore, two specific application examples of water which, however, gives rise to additional security vulnerabili-
supply distribution systems and wide-area power systems are ties [35], [93]. Specifically, 1) the security threats of CPSs are
presented to demonstrate the applicability of secure state estima- more severe in comparison with traditional networked con-
tion and control approaches. Finally, several challenging issues trol systems because of their broader and extendable attack
are discussed to direct future research. surfaces; 2) adversaries can lead to significant disruption of
Index Terms—Cyberattacks, cyber-physical systems (CPSs), physical processes by means of malicious operations on dif-
power systems, secure control, secure state estimation, water ferent cyber hierarchies [73]; and 3) any successful attack on
distribution systems. CPSs could point to catastrophic system failure and result in
unaffordable losses, which in turn serves as a profit-oriented
motive for real-world adversaries. These facts are further
I. C YBER -P HYSICAL S YSTEMS testified by several security incidents of modern infrastruc-
HE EFFICIENT integration of physical processes
T and cyber infrastructure with ubiquitous computation
resources and communication capabilities greatly boost the
ture, such as Ukrainian power grid attacks, Stuxnet worm
attacks on Iran’s nuclear control systems, distributed denial-
of-service (DDoS) attacks on Internet infrastructure company
rapid development of cyber-physical systems (CPSs) in both Dyn, see [24], [64], [85] for more examples. These secu-
rity incidents testify that attacks on CPSs, especially those
Manuscript received November 8, 2020; accepted November 24, 2020. Date launched in the cyber domain, can lead to considerable losses
of publication December 22, 2020; date of current version January 12, 2021.
This work was supported in part by the Australian Research Council Discovery in the economy and even human lives. Therefore, the security
Early Career Researcher Award under Grant DE200101128. This article of CPSs via developing new analysis techniques and design
was recommended by Associate Editor O. Kaynak. (Corresponding authors: countermeasures is becoming more significant than ever in
Jun Wang; Qing-Long Han.)
Derui Ding, Qing-Long Han, and Xiaohua Ge are with the School both academia and industry.
of Software and Electrical Engineering, Swinburne University of Secure state estimation and control constitute an effective
Technology, Melbourne, VIC 3122, Australia (e-mail: [email protected]; and promising means for addressing various security-related
[email protected]; [email protected]).
Jun Wang is with the Department of Computer Science, City University issues of CPSs. The main objective is to keep an acceptable
of Hong Kong, Hong Kong, and also with the School of Data Science, City performance level of the CPS by resorting to different security
University of Hong Kong, Hong Kong (e-mail: [email protected]). countermeasures, including attack attenuation and mitigation,
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TSMC.2020.3041121. isolation, detection, and compensation. When an attack occurs,
Digital Object Identifier 10.1109/TSMC.2020.3041121 the developed secure estimation/control mechanisms possess
2168-2216 
c 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: University of Canberra. Downloaded on May 20,2021 at 11:45:52 UTC from IEEE Xplore. Restrictions apply.
DING et al.: SECURE STATE ESTIMATION AND CONTROL OF CPSs: A SURVEY 177

certain capabilities to mitigate or counteract attack effects, or evaluation indicators and attack defense strategies. Then, the
prevent CPSs from severe performance degradation and loss, recent secure control results are discussed and classified into
or allow the system designers to make corrections and recover different categories. Furthermore, two typical application sce-
the system from any unsafe operation. Different from vari- narios are exemplified to demonstrate the applicability of the
ous network-induced phenomena, such as data packet dropouts secure state estimation and control approaches. Finally, sev-
and channel fading, the statistical characteristics or structured eral challenging issues are suggested to direct future research
information of cyberattacks may be unknown. Therefore, some investigations.
existing analysis and design approaches for networked control
systems [37], [98] must be revaluated in CPSs. Furthermore,
research and development of novel secure estimation and con- II. T YPICAL C YBERATTACKS
trol are demanded to systematically understand the interactions Cyberattacks on CPSs can be roughly classified into two
of the physical processes with the cyber elements for guaran- types: 1) denial-of-service (DoS) attacks and 2) deception
teeing better system performance, resilience, and security of attacks, depending on the adversary’s security goals of the data
CPSs subject to cyberattacks. exchanged through communication networks. Specifically,
In the past few years, there are several survey papers of DoS attacks aim to disrupt data availability and exchange-
security-oriented CPSs in [14], [20], [46], [64], [72], and [85]. ability by maliciously consuming communication/computation
For example, the recent progress of secure communication and resources, such as occupying channels or consuming CPU
control of smart grids under malicious cyberattacks is reviewed or memory resources. As a result, data may not be reached
in [72], where different attack models and effects as well as or sent to its destination successfully or completely, and the
security strategies are reviewed from IT protection and secure value field of mapping function from a sender and receiver
control-theoretic perspectives. A summary of detection meth- can be denoted as a null set if data availability is com-
ods of false data injection (FDI) attacks on smart grids is pletely disrupted. Deception attacks, on the other hand, intend
made in [64]. The existing FDI attack detection algorithms to compromise data integrity and trustworthiness by manip-
in smart grids are classified into model-based types and data- ulating packets over communication networks while keeping
driven types. From a systems and control perspective, the CPS stealthy to detectors. Some typical deception attacks include
security issue is evaluated in [14], where some latest systems FDI attacks, replay attacks, covert attacks and zero-dynamics
and control methods are reviewed and classified into preven- attacks. A cyber-physical attack space is established in [82]
tion, resilience, and detection and isolation. An overview of and [83] to provide a comprehensive taxonomy of exist-
security control and attack detection for industrial CPSs is con- ing cyberattacks based on CPS model knowledge, disruption
ducted in [20]. An intensive discussion of adversarial attacks resources, and disclosure resources.
and their defenses is provided in [46] for sensor-based CPSs in For the sake of performance analysis, the dynamic behav-
the field of computer vision. Emerging techniques improving ior of random DoS attacks can be evaluated by resorting to
the safety and security of CPSs and Internet-of-Things (IoT) Bernoulli sequences or Markov processes, where the latter has
systems are surveyed in [85] from two aspects: 1) design time the capability of disclosing channel recovery. Following this
techniques verifying properties of subsystems and 2) runtime idea, the security performance analysis is reliant on a pri-
mechanisms helpful against both failures and attacks. ori statistical information of the random DoS attacks, which
Profound security analysis and synthesis of CPSs rely on inevitably limits their applications in practical engineering as
suitable dynamical models of CPSs and reasonable mathe- it is generally nontrivial to identify the specific statistical
matical descriptions of cyberattacks. However, the inherently information of the attack signals. Furthermore, the motiva-
unknown attack knowledge and generally complex system tion for sophisticated adversaries adopting such probabilistic
dynamics induce several major roadblocks during security attack models requires further justification in realistic scenar-
performance analysis and countermeasure design of CPSs, ios [28], [73]. To circumvent such a limitation, a general DoS
including 1) complicated time series of data transmissions attack model is developed in [73] by posing limitations on the
subject to simultaneous cyberattacks and network-induced attack occurrence frequency and duration. CPSs under such
delays/data losses/packet disorders under various communi- DoS attacks are then modeled as a class of switched systems
cation protocols; 2) effective integration of active detection consisting of a closed-loop subsystem in an attack-free case
algorithms of cyberattacks and classical systems and control and an open-loop subsystem in an attack-present case. As such,
theories; 3) epidemic propagation of attack-incurred impacts the performance indicator such as average dwell time in con-
across physical and cyber domains; and 4) time-varying ventional switched systems can be employed to examine the
network topologies due to essential “plug-and-play” and scal- effects of such DoS attacks on the resulting estimation error
able functions. To the best of the authors’ knowledge, there system [99].
is a paucity of a detailed survey of the state-of-the-art results Replay attacks generally do not require preliminary model
in secure state estimation and control of CPSs. Therefore, it knowledge and can be initiated by following two key steps. In
is significant to conduct a comprehensive review of the recent the first step, some disclosure attacks are launched to record
advances of secure state estimation and control for CPSs and sensor measurements or control signals for some prescribed
provide an up-to-date evaluation of the conditions and limita- time. In the second step, the recorded historical data is
tions of existing results. Specifically, the latest results of secure retransmitted and replayed to estimators, filters, controllers, or
state estimation are reviewed in light of various performance actuators in such a way as to deteriorate system performance

Authorized licensed use limited to: University of Canberra. Downloaded on May 20,2021 at 11:45:52 UTC from IEEE Xplore. Restrictions apply.
178 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO. 1, JANUARY 2021

TABLE I
S TATE E STIMATION U NDER C YBERATTACKS as far as possible. In doing so, an indispensable assump-
tion is that the structured information of cyberattacks, such
as statistical information or boundedness information, is a
priori known. Following this idea, the main focus is then
placed on disclosing or offsetting the undesirable impact from
compromised data generated by malicious attacks. For exam-
ple, an algorithm of variance-constrained filtering over sensor
networks is proposed in [55] for discrete time-varying stochas-
tic systems whose physical constraint of bounded deception
attacks is described by two diagonal matrices with elements
unknown but involved over an interval. In [16], a distributed
or even jeopardize stability due to outdated information of the filter with double gains is designed, which can be regarded as
replayed data. Notice that the replaying start time and duration two weight matrices reflecting the different confidence levels
could be completely unknown at the receivers’ side, which of the information from itself and from neighboring nodes.
poses a significant challenge for security performance anal- The filtering performance analysis of double-gain Kalman-
ysis and countermeasure design. The traditional performance type filters is first performed, where a sufficient condition
analysis approaches on time-varying delays may offer some is derived to ensure the boundedness of the correspond-
inspiration to disclose the maximum allowable duration of ing covariance sequences with the help of the mathemati-
replay attacks on system performance in a worst-case scenario. cal induction approach. A state-saturated recursive filter is
The central aim of FDI attacks is to mislead the system proposed in [75] to account for both state saturations and
components, including estimators, filters, controllers, and actu- randomly occurring deception attacks in the framework of
ators, such that the system normal operation is disrupted. For complex networks, where attacks are adequately described by
this purpose, FDI attacks on CPSs can target both physical an attacked-measurement-output model. A recursive filtering
equipment and cyber infrastructure [64], and generally require algorithm is designed with variance constraints that can be
a certain amount of knowledge about physical and cyber implemented iteratively for online computation. It is shown
components, such as network topologies, physical connec- that the designed filters above with optimal gains can pas-
tion, and system parameter configurations, to generate stealthy sively restrain the destructive influences caused by tampered
attack sequences [13]. As a result, some well-structured attack data.
sequences can be designed to arbitrarily change system states Estimators or filters can also actively integrate some detec-
but not affect the residuals so as to escape any predeployed tion mechanisms to remove the compromised data generated
detector or monitor from triggering an alarm, leading to the by malicious attacks as much as possible. Benefiting from
so-called stealthy attacks or covert attacks. It is clear that their favorable statistical characteristics, χ 2 detector and its
stealthiness is of crucial importance to almost all realistic variants are widely adopted. For example, the adverse impact
adversaries. Such stealthiness can be also interpreted as dupli- of FDI attacks for wireless sensor networks is effectively
cating exactly the real unstable zero dynamics such that attack relieved by resorting to a protector, in which a randomly gen-
signals can gain a cover-up in the output nulling space, which erated threshold obeying a uniform distribution is employed
gives rise to zero-dynamics attacks [69]. On the other hand, it to preserve the Gaussianity of estimator dynamics. In light
should be noted that stealthy attacks are correctable via suit- of such a detection rule, a critical attack probability is ana-
able strategies if the systems are strongly observable [28], [80], lyzed [92]. It is shown that when the considered probability
and become detectable if only partial sensors are attacked [83] is bigger than some critical value, the steady-state solution
or some critical sensors are protected [4]. of estimation error covariance could exceed a preset value.
Recently, inspired by the Pauta criterion (i.e., 3δ criterion),
III. S ECURE S TATE E STIMATION OF CPS S χ 2 distribution-based detector is developed [15] to remove
abnormal data. Distributed filters are then designed such that
For capturing the real-time dynamics of CPSs, it is cru- their gains are captured via the recursive solution of a matrix
cial to reconstruct system states based on only measured yet equation on the upper bound of covariance of filtering errors.
possibly corrupted information from sensors. Different from Furthermore, the inequality of arithmetic means is employed
traditional control systems, the tight integration of physical and to deal with the challenge caused by cross-covariance matri-
cyber components, and the occurrence of various malicious ces, aiming at realizing a distributed implementation of the
attacks pose nontrivial challenges to the performance analysis developed algorithm.
and the design of state estimators or filters. In what follows, In practice, it is not uncommon that only a subset of
we classify some existing secure state estimation approaches sensors can be arbitrarily tampered with by attackers due
according to performance indicators and defense strategies mainly to their limited energy or disruption and stealthiness
against cyberattacks, as briefly summarized in Table I. resources. It is worth noting that the estimation performance
can be properly warranted if the corrupted sensor is accurately
A. Variance-Based Secure State Estimation detected and effectively isolated. For example, in light of a
The variance-based state estimation aims to select appro- sparse observability condition, an attack detection algorithm is
priate gain parameters to minimize estimation error variance designed [58] with the help of the sample average value and

Authorized licensed use limited to: University of Canberra. Downloaded on May 20,2021 at 11:45:52 UTC from IEEE Xplore. Restrictions apply.
DING et al.: SECURE STATE ESTIMATION AND CONTROL OF CPSs: A SURVEY 179

the attack-free expected value of the considered CPS. When constraining attacker capability and mitigating the impul-
the attack subset is properly identified, the performance of the sive and/or outlier-like effects of cyberattacks. Specifically,
developed algorithm does not exceed the one by the worst-case the anomaly on system measurement outputs caused by this
Kalman estimation. The optimal secure estimation is pursued type of cyberattack can be deemed as impulsive perturba-
in [77] for attacks without restrictions on their statistical prop- tions and/or measurement outliers. In this sense, the effects of
erties, boundedness, and time evolution in comparison with these attack-incurred abnormal measurements can be attenu-
the sparse attacks whose signals are bounded nonstochastic. ated effectively by constructing estimators with some saturated
A satisfiability modulo theory approach is then explored to output rejection. From this perspective, the saturation level can
overcome the scalability issue under brute force search. be explored to reveal the attenuation ability of the designed
state estimators. For example, a saturated innovation update
scheme is adopted in [12] for distributed state estimators with
B. Stability-Based Secure State Estimation an adaptive threshold of the saturation level, and in [79] for
Disturbance attenuation analysis provides an alternative stochastic nonlinear systems with a sector bounded condi-
framework for optimizing system performance in the pres- tion on the saturation constraint. It is shown in [12] that the
ence of malicious attacks. Secure filtering issues under this state estimate from any agent converges to its true param-
framework are investigated in [86] for positive systems in a eter, and the developed distributed algorithm is resilient to
distributed way, in [41] for complex networks with partial- sensor attacks and independent of the communication topol-
nodes-based estimators, and in [50] for linear discrete-time ogy as long as the agent network remains connected. In [88],
systems whose communication is scheduled by Round-Robin a saturated innovation scheme with an adaptive gain coeffi-
protocols. A common feature of those results is that the fil- cient and a mode switch mechanism is developed, where the
ter gains depending on the solution of some linear matrix mismatched unknown inputs are suppressed by resorting to the
inequalities are designed in response to the worst-case sce- well-known L2 -gain attenuation property. Dynamic saturations
narios of malicious attacks. In addition, for constrained DoS with an adaptive rule are further developed in [2] and [9]. It
attacks in terms of their occurrence frequency and duration, is noted that dynamic saturations with adaptive saturation lev-
some sufficient conditions under two nonlinear constraints els enjoy more flexible attack attenuation capability and less
on the predetermined frequency and duration parameters are estimation performance degradation.
derived in [56] to evaluate whether the resultant filtering error Ellipsoidal and maximum correntropy indicators are preva-
system satisfies the prescribed H∞ disturbance attenuation lently employed to evaluate the security and estimation
performance. It is further noted that the switching relation- performance for CPSs. For example, the weighted maxi-
ship between stable and unstable subsystems can be adequately mum correntropy criterion exhibits the potential capability
exploited to derive the tolerable range of attack frequency and of keeping appropriate distribution characteristics and mean-
duration. Besides, the stability condition in the mean square while restraining abnormal data due to cyberattacks. On
sense can be adopted to capture the tolerable attack probability. the other hand, cyberattacks on sensor measurements could
Attack isolation and attack attenuation from the active result in some significant deviations between the predicted
defense perspective represent two typical approaches to coun- ellipsoidal set and the updated ellipsoidal set during attack
teract the negative effects of malicious attacks. In the past occurrence. Such a feature can be explored to detect and
few years, attack identification and state estimation of CPSs isolate the attacks, where some a priori assumptions of
subject to sparse attacks have been attracting ever-increasing attacks, such as their statistics, magnitude, occurred place and
research interest. For example, a fundamental limitation (i.e., time, are not required. For example, the distributed filtering
the necessary condition of 2κ-observability which means that issue is investigated in [81] for general stochastic nonlinear
the system still remains observable when removing any 2κ systems subject to deception attacks. The developed scheme
sensors) is disclosed for noiseless systems [28], [78] and is related to only local information of each node, thereby
noisy systems [67], [68]. Such a limitation is relaxed to the satisfying the requirement of distributed implementation. A
2κ-detectability [6], [65]. Furthermore, the secure state esti- set-based detection rule, which is performed by checking the
mation problem is transformed into the solvability of an 0 intersection of the two ellipsoids of state prediction and state
optimization issue in [28] and an 1 /r optimization issue estimate, is presented in [36] and [63] to monitor attack
in [47], or the performance analysis problem of 2 , H2 , and occurrence in a timely manner. Resilient and secure mon-
H∞ systems in [65] by virtue of the classical robust control, itoring is further demonstrated by retrieving the previous
and fault detection and isolation methods. Meanwhile, secure secure measurement and giving up the estimator update when
state observers are designed for CPSs subject to sparse sen- an attack occurs. On the other hand, the switched gradient
sor attacks of different attack modes via boundedness analysis descent technique and the modulo theory approach, as well
of input-to-error outputs [6]. An adaptive switching scheme as the set cover approach under the observability assump-
is performed to confirm the correct attack mode in the actual tion, are, respectively, presented in [52], [53], and [77] to
running stages. It is shown that the attack detector has the handle the intrinsic combinatorial complexity issues during
capability of automatically reaching and remaining in the secure state estimation in [28] and [47]. The results above
appropriate mode and then turning off the attacked channels. further show that secure state estimation can be pursued with
Employing some artificial saturation constraint on state preserved scalability, convex relaxation, or convergence rate
estimators is regarded as a promising security measure for requirement.

Authorized licensed use limited to: University of Canberra. Downloaded on May 20,2021 at 11:45:52 UTC from IEEE Xplore. Restrictions apply.
180 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO. 1, JANUARY 2021

TABLE II
S ECURE C ONTROL U NDER C YBERATTACKS attack frequency and duration, is put forward in [73] for a
linear system with a static state feedback controller. Such a
robustness measure is further extended in [29] by resorting to
an impulsive controller based on a dynamic observer. A cyclic
dwell-time switching strategy is proposed in [102] to describe
a class of periodic DoS attacks such that the induced closed-
loop system is transmitted into a cyclic switched system. Based
on such a transformation, conditions ensuring the predeter-
mined H∞ performance are derived by means of modal dwell
time corresponding to the stable subsystem and the unstable
one. Furthermore, a switching signal taking values in a finite
set is employed to model the number of consecutive DoS
attacks in [74], where the corresponding stability criterion is
derived by making use of a switching parameter-dependent
Lyapunov function.
IV. S ECURE C ONTROL OF CPS S Active detection of cyberattacks offers an effective means
to enhance the system’s adaptation to malicious attacks. There
There are two substantial lines of research on secure control is no doubt that making existing attacks detected and removed
for CPSs under cyberattacks, which are categorized as attack- largely promotes the system security. In the context of secure
tolerant control and attack-compensated control. The first control, it should be mentioned that the existing detection
category focuses on the design of a suitable control policy/law schemes, such as χ 2 detectors and cumulative sum detectors,
to tolerate unpredictable anomalies caused by attacks [100]. A rely on monitoring the changes of the evaluated cost func-
resilience concept is fashionably put forward to characterize tions under different attack assumptions, or the gaps between
the tolerant degree of the resulting closed-loop system stability the real sensor measurements and the predicted measurements.
and performance against attacks. An emphasis is then placed For example, an adaptive switching logic is exploited in [5]
on examining the prescribed tolerance capability or pursuing to provide an online location of the real system mode via
the maximal tolerance capability for the controlled system, observing the variation of the traditional quadratic cost in
allowing further intervention actions to be made from the the framework of linear quadratic control. A Kalman-based
system designers. The second category, on the other hand, is attack detector with an observation window of a given length
concerned with the design of preferable compensation schemes is designed in [26] to remove the occurred deception attacks.
to prevent the system performance and stability from get- A common feature of the above detectors is that the dura-
ting severely deteriorative or even becoming unstable. For this tion of deception attacks is captured to describe their negative
purpose, it is indispensable to implement appropriate attack effects. Then, the maximum allowable duration threshold is
detection mechanisms to identify and locate the occurrence examined to maintain the desired system stability. In [45],
of cyberattacks. A summary of some existing secure control the probability of packet losses for the channel modeled by
results is provided in Table II. In what follows, the latest the signal-to-interference-plus-noise ratio is dependent on the
development of secure control is evaluated from three aspects: channel gain which is determined by the energy of jamming
1) centralized secure control; 2) distributed secure control; and attackers. It is shown that the control cost is not monotonic
3) resource-aware secure control. with respect to the channel gain, and the effect is considerable
for the mismatch between the nominal channel gain and the
A. Centralized Secure Control true one.
A direct consequence of DoS attacks on CPSs is that the Defenders and attackers essentially play a noncooperative
system is operated in an open-loop manner as the desired con- game, which contributes to the development of game theory
troller is not capable to receive any sensor data for feedback. In in addressing various secure control problems of CPSs [94],
this case, the celebrated switched system theory finds its way [97]. The derived saddle-point equilibrium, from the perspec-
into ensuring secure control for CPSs under such DoS attacks, tive of attackers, reflects the lowest attack intensity such that
where the system is closed-loop in attack-free case but acts in the system under the admissible control strategy is driven out
open-loop otherwise. It is noteworthy, however, that the result- of the security zone. Such an intensity discloses the maxi-
ing system performance depends on the running duty cycle, mum tolerance capability of cyberattacks. Following this idea,
which is commonly known as dwell time, between the two an optimal control strategy involving both cyber defenders
cases. As such, the primary goal of secure control is to find the and DoS attackers is well explored [94], [95]. Specifically,
tolerant duration and/or attack frequency such that the desired an inverse game approach is proposed in [94] to compen-
system performance remains achievable. In other words, the sate for the attack-induced performance loss on the side of
secure control objective follows the traditional pattern of con- defenders and a dichotomy algorithm is presented to obtain
trol procedures based on pulse-width modulation, aiming to the lowest attack intensity. Recently, different optimal strate-
seek the permitted maximum on/off cycle to guarantee reli- gies by means of a zero-sum stochastic game approach are
able control performance. For example, a robustness measure derived in [57] for CPSs under replay attacks and FDI attacks,
against DoS attacks, which describes the tolerable maximum with an aim to strike a balance between control cost and

Authorized licensed use limited to: University of Canberra. Downloaded on May 20,2021 at 11:45:52 UTC from IEEE Xplore. Restrictions apply.
DING et al.: SECURE STATE ESTIMATION AND CONTROL OF CPSs: A SURVEY 181

security requirement. In [91], an intrusion detection system secure control issue of CPSs with an understanding of similar-
is deployed on the control side to defend against cyberattacks, ities of both mathematical descriptions and practical influences
where the expectant DoS frequency and duration are depicted between faults and certain cyberattacks. For example, a dis-
as some cost function consisting of frequency lower-bound tributed state predictor is employed to estimate the existing
and duration upper-bound. attacks, and then a resilient controller is designed to guaran-
Security of a practical CPS is regarded as a hard constraint tee robust performance and to adaptively compensate for the
that describes a guard line guaranteeing economic benefit and influence of attacks [61], [62].
life safety. However, it is nontrivial to achieve absolute security
for CPSs suffering from various stochastic disturbances and
randomly occurring cyberattacks. Despite a few attempts [40], C. Resource-Aware Secure Control
the security of CPSs in an almost sure sense is generally The efficient utilization of limited communication resources
difficult to be maintained from a control-oriented perspec- in CPSs stimulates extensive research interest from the control
tive. As a result, an alternative indicator, known as security realm. In the past few years, various communication schedul-
in probability, is exploited [21]. A framework of security ing protocols are employed to govern the token, with which
control in probability is established to disclose the transient the subsystems (or agents) are endowed the authority to access
dynamics of general nonlinear stochastic systems under decep- communication channels at each time instant. Typical schedul-
tion attacks. Furthermore, an original easy-solution scheme of ing protocols include, but are not limited to, event-triggered
desired controller gain is derived via the matrix inverse lemma. protocols [34], [37], [38], [98], random access protocol [103],
and round-robin protocols [10], [17], [18], [104]. Specifically,
it is well acknowledged that a well-designed event-triggered
B. Distributed Secure Control protocol leads to lesser samples or updates of filters/controllers
The physical components of practical CPSs could be than traditional time-triggered protocols, which greatly reduces
deployed in a spatially distributed way, which results in the unnecessary occupancy of the finite communication and
broader attack surfaces than traditional networked control computation resources. However, the system performance is
systems and further aggravated security risks. The system inevitably degraded to some extent because less sensor/control
performance could be also severely degraded due to corrupted data is adopted for achieving feedback control. In the context
sensor and control data during network communication. As of communication scheduling, it becomes apparent that cyber-
such, a critical concern is to design a suitable distributed secure attacks can result in a more tremendous data sparsity issue.
controller to render the resulting closed-loop system survivable This further leads to some inherent and nontrivial challenges
or recoverable from cyberattacks. for performance analysis and secure control design of CPSs
The overall performance of distributed CPSs consist- that are beyond the capacity of the existing results on stability
ing of multiple subsystems (or agents) is deeply related analysis and controller design of event-based control systems
to the information sharing among neighboring subsys- without cyberattacks.
tems. Distributed secure control that embeds attack model Note that the time series of data transmissions or updates
information (i.e., statistical or structured information) is under communication schedules become more complex due
proved to exhibit the capability of attack attenuation although mainly to the interference of malicious attacks, which poses
it is a passive control strategy [89]. For example, an impul- a significant challenge for continuous-time physical systems.
sive controller against randomly occurring deception attacks Under the assumption that the execution period and a uni-
is developed in [39] to realize the secure synchronization of form lower bound of sleeping periods are a priori known, a
leader–follower multiagent systems, where the synchronization sufficient condition of exponential stability is derived in [43]
errors lie in a bounded set dependent on the attack probability. by using the piecewise Lyapunov functional along with
In the presence of distributed DoS attacks, a control protocol a reconstructed state-error-dependent switched system. An
guaranteeing scalability and robustness is proposed in [90] for event-triggered scheduling and control co-design algorithm is
multiagent systems under event-triggered communication. developed in [71] to obtain both the triggering parameter and
For preseving network robustness, a resilient consensus the control gain. The quantitative relationship between the
issue of distributed CPSs is extensively investigated under κ allowable duration of DoS attacks and the triggering parameter
malicious and/or Byzantine attacks [1], [70], [84]. The effect governing the transmission burden is also revealed. For elimi-
of the trusted equipment (or nodes) is further explored in nating the impact from DoS attacks on time series, traditional
the past few years. For example, it is disclosed in [70] that event monitors based on measurement variations are replaced
κ Byzantine agents can be detected and identified by each by those based on system states. Such an event-triggered
agent in normal behavior when the linear consensus network is scheduling scheme is improved by integrating measurement
(2κ+1)-connected. A consensus algorithm is developed in [32] variations with a minimal trigger sleeping interval in order to
by removing the smallest and greatest κ values in absence of avoid the well-known Zeno behavior [44], [54]. The quanti-
trusted edges, and the resilient consensus can be achieved if tative relationship between the allowable attack duration and
the network is (2κ + 1)-robust with respect to a trusted edge the triggering parameter is disclosed in [54], where the event
set or a directed graph whose trusted edge set contains a span- is dependent on the measurement variations, the trigger sleep-
ning tree. On the other hand, the classical fault detection and ing interval, and the time instant determined by the observer.
estimation approaches provide a foundation to deal with the On the other hand, under a sparse observability condition,

Authorized licensed use limited to: University of Canberra. Downloaded on May 20,2021 at 11:45:52 UTC from IEEE Xplore. Restrictions apply.
182 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO. 1, JANUARY 2021

an observer in a delta domain is designed in [33] to esti-

mate the system state under sensor and actuator attacks, and
a self-triggered controller is designed via iterative analysis.
In the context of distributed secure control, there are con-
siderable results reported for CPSs under event-triggered com-
munication scheduling. For example, a sufficient condition is
derived in [19] to guarantee the desired consensus for a class of
stochastic multiagent systems subject to both lossy sensors and
deception attacks. A scalable design algorithm is established
by resorting to the second and the latest eigenvalues of the
Laplacian matrix. In [30], two elaborate interval classifications
are constructed by introducing the upper bound of adjacent
event intervals under DoS attacks, their duration and their
launching time, and then the switched system theory is per-
formed to derive the consensus condition. It should be noted
that the presence of cyberattacks makes the exclusion of Zeno
Fig. 1. SCADA water distribution system under attacks (xk =
behavior from the designed distributed event-triggered secure (1) (2) (3) (4) (1) (2) (3) (4)
[xk , xk , xk , xk ]T : pressure heads; yk = [yk , yk , yk , yk ]T : sensor
controllers generally difficult. This is because the interval of (1) (2) (1)
measurements; wk = [wk , wk ]T : users’ consumptions; uk = [uk , uk ]T :
(2)
two consecutive data transmissions may not be that of two control input flow rates; A1, A2: physical attacks on pump and reservoir/
adjacent events invoked. Furthermore, some existing active tank; A3: cyberattacks on sensor/control data).
attack detection and compensation approaches need to be re-
evaluated due to the undistinguishable data sparsity caused by
DoS attacks and event-triggered scheduling. of attack feasibility and their countermeasures via watermark-
From a system-wide perspective, attack detection and secure ing or noisy control are established to maintain optimized
control are required to be seamlessly integrated. This conse- detection probability of attacks at the cost of compromised
quentially leads to a tradeoff between detection precision and control performance. By virtue of statistical detection the-
real-time control performance. It is thus desirable to estab- ory, sequential detection and isolation methods are evaluated
lish a co-design approach for detection and control parameters in [24] and [25] for detecting transient changes caused by
in such a way as to achieve acceptable security performance cyber-physical attacks in stochastic SCADA water distribution
and desirable control performance, which deserves an in-depth systems.
investigation. Protecting critical infrastructure like a water distribution
system from malicious attacks is generally difficult. The
system is often made up of a range of physical elements,
coupled with integrated computerized and communication
V. T WO T YPICAL A PPLICATION S CENARIOS OF S ECURE
technology, connecting a range of end users, water facili-
S TATE E STIMATION AND C ONTROL
ties, plants, stations/substations, distribution and transmission
A. Secure State Estimation of Water Distribution Systems infrastructure. There is thus a clear need to quantify the
Supervisory control and data acquisition (SCADA) systems, impacts of adversarial attacks and evaluate the effectiveness
as a paradigm of CPSs, are broadly employed to provide of countermeasures.
remote monitoring and control solutions for geographically In what follows, we outline a secure set-membership state
dispersed assets, including electrical power grids, water, oil estimation framework for a SCADA water distribution system,
and gas pipelines, and sewage treatment plants. The reli- which provides a secure estimation of the unavailable system
able operation and protection of modern critical infrastructure, state in the presence of unknown disturbance input, measure-
including water distribution systems rely on secure state esti- ment noise, and malicious attacks.
mation mechanisms, as testified by the vast literature on A SCADA water supply distribution system typically
secure state estimation for SCADA systems. For example, involves three layers, as briefly depicted in Fig. 1: 1) a physi-
the problem of security threat assessment for a SCADA irri- cal layer, where numerous waste-water treatment plants, water
gation canal system under remote water pilfering attacks is storage reservoirs, tanks, pipelines, sensors (e.g., pressure,
considered in [3]. A field operational test is performed to flow, and level sensors), junctions, actuators (e.g., pumps and
demonstrate that such attacks can be stealthy and bypass valves) and other hydraulic devices are employed to offer
detection by the SCADA system without being detected. The water transmission and distribution to end users; 2) a cyber
problems of detection and isolation of attacks on a water distri- layer, where a remote SCADA control center, consisting of
bution network of cascaded water canal systems are addressed state estimators, anomaly detectors, controllers, communica-
in [4]. A model-based fault detection and isolation scheme tion devices and other functions, is established for real-time
based on a bank of unknown input observers is implemented monitoring and supervising control; and 3) a network layer,
to detect and isolate the attack signals in canal. In [59], which enables two-way data transmissions among smart sen-
the model-based techniques are developed to detect replaying sors, remote control center and actuators. We employ the
attacks on the sensors of a SCADA system. The conditions following discrete-time state-space model for describing the

Authorized licensed use limited to: University of Canberra. Downloaded on May 20,2021 at 11:45:52 UTC from IEEE Xplore. Restrictions apply.
DING et al.: SECURE STATE ESTIMATION AND CONTROL OF CPSs: A SURVEY 183

 
SCADA monitoring and control system [24], [59]: −Pk+1 k
s.t. T ≤0 (4)
 k k
xk+1 = Axk + Bu uk + Bw wk + Ba ak
yk = Cxk + Dw wk + Dv vk + Da ak
(1) at each time step k ∈ {0, 1, . . . , Tf } to obtain the
real-valued matrix sequences Pk+1 > 0, Gk , Lk , and
where xk ∈ Rnx denotes the system state; uk ∈ Rnu denotes scalar sequences θm,k > 0, m = 1, 2, . . . , 6, where
the control input; wk ∈ Rnw represents the general disturbance k = [(A − Gk −Lk CA)x̂k − Lk CBu uk , AEk − Lk CAEk ,
input (including process noise and system faults); yk ∈ Rny is Bw −Lk CBw , Ba − Lk CBa , −Lk Dw , −Lk Dv , −Lk Da ], k =
the actually received sensor measurement output at the remote diag{ 6s=1 θs,k − 1, −θ1,k I, −θ2,k Q−1 −1 −1
k , −θ3,k Sk , −θ4,k Qk+1 ,
control center; vk ∈ Rnv represents the measurement noise; −θ5,k R−1 −1
k+1 , −θ6,k Sk+1 }, and Ek is recursively obtained from
ak ∈ Rna denotes the attack signal at time k satisfying Pk = Ek Ek by a Cholesky factorization.
T
 Step 2: Design the following state estimator based on the
φk−k0 +1 , k ∈ TL received sensor measurement yk+1 and gain matrices Gk , Lk :
ak = (2)
0, k  ∈ TL
x̂k+1 = Gk x̂k + Bu uk + Lk yk+1 (5)
with TL = {k0 , k0 + 1, . . . , k0 + L − 1} being the attack dura-
tion and FL = {φ1 , φ2 , . . . , φL } being the attack profiles. where x̂k+1 ∈ Rnx denotes the state estimate at time k + 1.
Note that the attack start time k0 , the time length L, and the Step 3: Derive the state estimate set {xk+1 : xk+1 = x̂k+1 +
nonzero vectors φ1 , φ2 , . . . , φL ∈ Rna are unknown to system Ek+1 α, α ∈ Rnx , α ≤ 1} based on x̂k+1 and Pk+1 , which
designers and can be carefully conditioned by adversaries to encloses all possible values of the true system state xk+1 and
cause malicious consequences; A, Bu , Bw , Ba , C, Dw , Dv , and ensures that (xk+1 − x̂k+1 )T P−1 k+1 (xk+1 − x̂k+1 ) ≤ 1 for any
Da are real-valued matrices of appropriate dimensions. The k ∈ {0, 1, . . . , Tf }.
system is considered to be in normal operation if ak ≡ 0 for The secure state estimation problem is cast into the feasi-
all k ∈ N. Without requiring specific statistical properties, it bility of a recursive convex optimization problem (3) subject
is assumed that the disturbance, noise and attack inputs are to a series of linear matrix inequalities in (4). A detailed
unknown but bounded (UBB) and satisfy wk ∈ E k (0, Qk ), derivation of (4) can be found in [36]. Note that the com-
vk ∈ E k (0, Rk ) and ak ∈ E k (0, Sk ), where E k (ck , k )  puted state estimate in step 3 is inherently a set in state-space
{zk : (zk −ck )T −1
k (zk −ck ) ≤ 1} represents an ellipsoid enclos- at each time step k + 1. Therefore, the derived ellipsoidal set
ing a real vector zk ∈ Rn with a real vector ck ∈ Rn being E k (x̂k+1 , Pk+1 ) provides a guarantee that all possible values of
the center and a real-valued matrix k = Tk > 0 being the the SCADA system’s true state xk+1 consistent with the UBB
shape matrix of the ellipsoid, respectively. The system’s initial disturbance, noise and attacks reside in the ellipsoid, which
state x0 is supposed to reside in a given ellipsoid E 0 (x̂0 , P0 ) is distinct from the traditional estimation methods, such as
with x̂0 denoting the initial state estimate and P0 = PT0 > 0 H∞ estimation and Kalman filtering since they only generate
denoting a given matrix. a single vector estimate at each time step. Apart from using
y
The vector ak is decomposed as ak = [axk ; auk ; ak ], where ellipsoids to bound the state estimates, alternative set represen-
y
axk ∈ Rnx , auk ∈ Rnu , and ak ∈ Rny denote the attack vectors tations, such as interval, parallelotopes, and zonotopes, have
affecting system state, control signal, and sensor measure- also been widely employed to assess the bounding estimation
ments, respectively. Accordingly, the matrices Ba and Da in (1) performance of various dynamical systems in the presence of
y
are decomposed as Ba = [Bxa , Bua , 0] and Da = [Dxa , Dua , Da ]. It UBB uncertainties. Although it is generally difficult to evalu-
is stressed that a sophisticated adversary is capable of exploit- ate different set-membership estimation methods in a unified
ing various attack strategies, such as gaining physical access, framework, estimation accuracy and computational efficiency
jamming, falsifying, replaying, and a combination of them, are the two key requirements worthy of careful consideration
without getting detected, which is an obvious threat to SCADA and negotiation.
systems. Consider a SCADA water distribution system shown in
The secure state estimation problem of the SCADA Fig. 1 whose system parameters are borrowed from [25].
system (1) is formuated as follows. For any wk ∈ E k (0, Qk ), During normal operation, reservoirs R1 and R2 are sup-
vk ∈ E k (0, Rk ), and ak ∈ E k (0, Sk ), the objective is to design a (1)
plied with constant flow rates uk = 0.4533 m3 /s and
bounding ellipsoidal set E k (x̂k+1 , Pk+1 ) at each time step k+1 (2)
uk = 0.5539 m /s, respectively. The two end users’ demands
3
which guarantees to enclose the SCADA system’s true state both fluctuate between 0.9 m3 /s and 1.1 m3 /s. Sensors S1–
xk+1 in (1) regardless of the simultaneous presence of UBB S4 suffer from random measurement noises causing variations
disturbance wk , measurement noise vk , and malicious attacks of (0.05, 0.25) m on their readings. Set Qk = 2I, Rk = I,
ak , where the ellipsoid center is specified by the desired state Sk = 30I, x0 = [100; 80; 30] m, x̂0 = [98; 82; 29] m, and P0 =
estimate x̂k+1 and the shape matrix is Pk+1 = PTk+1 > 0 to be 50I. Furthermore, let Bxa = I, Bua = [0.6250, 0; 0, 0.833; 0, 0],
determined. y
Dxa = 0, Dua = 0, and Da = I. We next evaluate the secure
The following three steps outline the design procedure for estimation of the SCADA system under various attacks.
the above secure state estimation problem. Case I (FDI and DoS Attacks): We first consider that an
Step 1: Solve the following convex optimization problem: adversary deliberately alters the measurements of all sensors
via concurrent FDI and DoS attacks (attack point A3 in Fig. 1).
min trace(Pk+1 ) (3)
Pk+1 >0,θm,k >0,m=1,2,...,6 The FDI attacks inject some erroneous/misleading information

Authorized licensed use limited to: University of Canberra. Downloaded on May 20,2021 at 11:45:52 UTC from IEEE Xplore. Restrictions apply.
184 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO. 1, JANUARY 2021

into sensor data over the communication network. Specifically,

the FDI attacks corrupt all sensor measurements by adding
y
ak = [r1 ; r2 ; r3 ; r4 ] with ri , i = 1, 2, 3, 4 being some ran-
dom values on [−2, 2], occurring at k0 = 10 and lasting for
L = 10 time steps. The DoS attacks adopt a jamming strategy
for preventing the measurements from reaching the remote
control center. A direct consequence of DoS attacks is that
the measured outputs recorded by sensors are partially lost
during their network transmissions. We assume that the DoS
attacks cause up to 5% missing of the normal measurement
yk [without ak in (1)] at the receiver side from k0 = 50, i.e.,
y
ak = r5 yk , where r5 is a random variable ranging on [0, 0.05].
Furthermore, axk = [0; 0; 0] and auk = [0; 0] in this case.
Case II (Stealthy Attacks): An ultimate objective of attack
on a SCADA water distribution system is to disrupt the nor-
mal water distribution without being monitored or detected.
We next evaluate a stealthy attack case, where the adversary
performs some physical attacks on water reservoir and pump,
causing the reservoir water levels to decrease significantly,
while at the same time implementing some cyberattack strat-
egy, such as FDI attacks to cover up the changes within the
recorded sensor measurements, rendering the attacks stealthy
to the remote monitor/detector. Specifically, it is assumed
that ∀ k ∈ TL = {30, 31, . . . , 39}, the attack profiles are
taken as axk = [0; −0.5; 0], auk = [−0.4533; 0] and xk+1 a =
Axk + Ba ak + Bk ak , for any {xk }k≤30 = 0, k ∈ {30, 31, . . . , 39},
a x x u u a
Fig. 2. Secure state estimation of the SCADA water distribution system
y (1) (2) (3)
and ak = −Cxka , for any k ∈ {30, 31, . . . , 39}, where xka under attacks. (a) System state xk = [xk , xk , xk ]T and estimate x̂k =
(1) (2) (3)
denotes the system state component induced by the attack. [x̂k , x̂k , x̂k ]T in case I. (b) Quadratic estimation error (xk − x̂k )T P−1 k (xk −
Furthermore, the first nonzero element in axk implies a physi- x̂k ) in case I. (c) xk and x̂k in case II. (d) (xk − x̂k )T P−1
k (xk − x̂k ) in case II. (e)
cal “water-pilfering” attack on R1 (attack point A2 in Fig. 1) xk and x̂k in case II under protected sensors S1 and S2. (f) (xk − x̂k )T P−1 k (xk −
x̂k ) in case II under protected sensors S1 and S2.
with a constant water pilfering flow rate 0.4 m3 /s. The second
nonzero element in auk indicates a physical pump-off attack
on P2 (attack point A1 in Fig. 1). To keep the changes on
hand, when the channels of the critical sensors S1 and S2 are
the pressure heads of R1 and R2 unavailable to the remote
y protected (e.g., via encryption) but the channel of S4 remains
monitor/detector, a further FDI attack ak is performed via alter-
y attacked, the state estimator offers accurate and prompt obser-
ing the sensor data with the misleading information ak such
vations of the true system state, as depicted in Fig. 2(e) and (f).
that the state estimator cannot identify any change caused by
This further implies that these attacks become not stealthy
the physical attacks on the received measurements during the
when only a subset of the sensor data channels is attacked [83].
attack period.
In general, whether or not attacks are stealthy depends on
The simulation results are demonstrated in Fig. 2. It can
the adversaries’ system model knowledge as well as their
be seen that: 1) in case I, the state estimate ellipsoid guar-
disclosure and disruption resource. Furthermore, securing crit-
antees the enclosing of the true system state xk+1 regardless
ical sensors against strategic manipulation could lead to better
of the disturbance and noise as well as the simulated cyber
estimation and detection performance against stealthy attacks,
FDI and DoS attacks on the sensor data. This also shows
which serves as a reasonable cyber-defense strategy [4].
that simple cyberattack strategies via corrupting data avail-
ability and integrity may not cause sufficient disruption of the
CPS and 2) in case II, although the true system state xk+1 B. Security Control of Wide-Area Power Systems
resides in the state estimate ellipsoid, the pressure heads at Modernized smart grids as well as traditional power
R1 and R2 decrease notably from k = 30 since the water was systems, as a typical representative of CPSs, are more vulner-
being withdrawn from R1 and the pump P2 was turned off. able to various malicious attacks because of their distributed
However, the state estimator cannot track accurately the true deployment and the utilization of various networked equip-
system state because of the concealed measurement changes ment [8], [93]. Their reliable operation is tightly related
on sensors S1–S4, and thus the attacks are stealthy to the to the system’s real-time information gathered by remote
remote monitor. This confirms that attacks, having access to terminal units and transmitted to the remote control center
all measurement transmission channels, can be stealthy and via some communication network. From the perspective of
bypass any anomaly detector [83]. In other words, via collud- the security life cycle, the deployed secure monitoring and
ing with some physical attacks, cyber FDI attacks can lead control systems mainly include four modules: 1) prevention;
to the destruction of the remote control center. On the other 2) detection; 3) mitigation; and 4) resilience. Up until now,

Authorized licensed use limited to: University of Canberra. Downloaded on May 20,2021 at 11:45:52 UTC from IEEE Xplore. Restrictions apply.
DING et al.: SECURE STATE ESTIMATION AND CONTROL OF CPSs: A SURVEY 185

significant detection algorithms for FDI attacks based on vari- known matrices determined by physical parameters of power
ous similarity tests are developed, where 1) the employed state systems. Here, θi,k , wi,k , Pmi,k , and Pvi,k represent,
estimators include weighted least squares estimators, median respectively, the deviations of the angular displacement of the
filters, maximum likelihood estimators, minimum mean square rotor frequency, rotating mass, generator mechanical output,
error estimators, Kalman filters, and unknown input observers and turbine valve position, see [11, Table I] and [48] in more
and 2) test indicators include Euclidean distance (L2 norm), detail. Ni stands for the neighboring set reflecting physical
largest normalized residual, chi-square test of residuals, the connections with area i.
cumulative sum of residual variations, and Kullback–Leibler Case I: The measurement output yi,k is subject to DoS
distances, as well as a combination of aforementioned indi- attacks and modeled as yri,k = θk yi,k . In this case, the
cators. The latest development of attack detection for power corresponding controller is designed as
systems is beyond the scope of this article and thus omitted.
ui,k = Kii yri,k = θk Kii yi,k
Resilient estimation and control of power systems modeled
by a dynamical state-space model are intensively studied. For where θk taking a value in {0, 1} describes DoS attacks, and Kii
example, the load frequency control issues are investigated is the desired controller gain. Over the time interval [k1 , k2 ),
in [71] for multiarea power systems subject to energy-limited the number and the duration of the launched DoS attacks
DoS attacks and in [48] for ones subject to hybrid cyberat- N (k1 , k2 ) and |T (k1 , k2 )| satisfy [73]
tacks, where co-design schemes are established to get both k2 − k1 k2 − k1
the parameter in event-triggered conditions and the controller N (k1 , k2 ) ≤ , |T (k1 , k2 )| ≤
Tf Td
gains. For general direct current microgrids facing the threat of
where positive constants Tf and Td meet Tf > 1 and Td > 1.
DoS attacks, some sufficient conditions quantified the system
By the switching system theory, the wide-area power
resilience are presented by means of convex optimization tech-
systems (6) is mean-square exponentially stable with a
niques in [49]. Furthermore, by virtue of Kalman filtering,
weighted l2 gain γ , if the following matrix inequalities hold:
the secure estimation issues of multiarea power systems are
addressed in [11] with a passive rejection of DoS attacks, AT0 P0 A0 − P0 ≤ αP0 (7)
and in [15] with active defense of FDI attacks, where the AT1 P1 A1 − P1 ≤ −βP1 (8)
established algorithms possess scalability benefiting from the
P0 ≤ π P1 , P1 ≤ π P0 (9)
solution scheme of cross-covariance matrices.
A hierarchical control framework of power systems, con- 3 ln(π ) < Tf (ln(ϑ) − ln(1 + β)) (10)
sisting of the primary control and the secondary con- 3(ln(1 + α) − ln(1 + β)) < Td (ln(ϑ) − ln(1 + β)) (11)
trol, is commonly established to ensure power quality and
where A0 = [Aij ]M×M and A1 = [Aij ]M×M + diag{Bii Kii Cii }M
frequency/voltage stability [22]. Specifically, the secondary
with the decision parameters, including two positive definite
control updates the set points of the local primary control via
matrices P0 and P1 , controller gain matrices Kii and posi-
sharing neighbor information. To mitigate an attack impact,
tive scalars α, β, π and ϑ ∈ (0, 1). The inequalities (7)
a finite-time control scheme is established in [23] to realize
and (8) describe the dynamical behavior of two switching
frequency regulation and active power-sharing. In [51], it is
subsystems, where the closed-loop subsystem provide addi-
shown that the islanded microgrid can maintain the desired
tional ability (i.e., parameter β) to stabilize the whole system,
stability with the help of designed mode-based secondary
and (10) and (11) reflect the constraint on the frequency
controllers. A small-signal model is employed to describe
and the duration of launched malicious attacks, respectively.
the microgrid as a jump system, and a vulnerability metric
Detailed deviation of the inequalities can be found in [56].
based on observability Gramians is defined to evaluate the
Case II: The measurement output yi,k (except i = 1) is
system security under random DoS attacks. In the presence
subject to deception attacks and modeled as yri,k = yi,k + ξi,k .
of unbounded injected attacks, resilient control frameworks
In this case, an attack-compensated controller is employed
are proposed in [105] to offer frequency regulation and volt-
age containment while guaranteeing the synchronization of ui,k = Kii yri,k − ηi,k
AC microgrids, and in [106] to preserve the proportional load where ηi,k , which can be regarded as the estimation of the
sharing and voltage regulation of DC microgrids. unknown attack information Kii ξi,k , is updated by
In what follows, we provide some preliminary results for  
controlling a wide-area power system under DoS attacks and ηi,k+1 = (1 − θ )ηi,k + θ Kii yri,k − Cii x̂i,k . (12)
deception attacks. The considered system including M areas is Here, x̂i,k is the estimation of xi,k at the time instant k, which
described by the following coupled discrete state-space model: is governed by the normal state predictor
  ⎧ 
xi,k+1 = Aii xi,k + j∈Ni Aij xj,k + Bii ui,k + Dii wi,k
(6) ⎨ Aii x̂i,k + j∈Ni Aij x̂j,k
⎪
yi,k = Cii xi,k + Eii νi,k
x̂i,k+1 = + Bii Kii Cii x̂i,k + Lii yri,k − Cii x̂i,k , i = 1
⎪
⎩ 
where xi,k = ( θi,k wi,k Pmi,k Pvi,k )T is the state of the Aii x̂i,k + j∈Ni Aij x̂j,k + Bii Kii Cii x̂i,k , i = 1.
ith area, yi,k stands for measurement outputs, which could
The wide-area power systems (6) is input-to-state stable if
be subject to cyberattacks, ui,k is the control inputs-based
the following matrix inequality holds
output-feedback, and wi,k and νi,k are energy-limited exter-
nal disturbance and noise. Aii , Aij , Bii , Cii , Dii , and Eii are ĀT P Ā − P < 0

Authorized licensed use limited to: University of Canberra. Downloaded on May 20,2021 at 11:45:52 UTC from IEEE Xplore. Restrictions apply.
186 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO. 1, JANUARY 2021

Fig. 3. Four area power system under attacks.

Fig. 5. Responses of the controlled power system with compensation ηi,k

under deception attacks.

Fig. 4. Responses of the controlled power system under DoS attacks: (a)
frequency deviation θ ; (b) tie-lie active power deviation w; (c) genera-
tor mechanical output deviation Pm ; and (d) valve position deviation Pv ,
where the gray rectangles depict the intervals occurred attacks.

with
B = diag{Bii }M ,  = diag{(1 − θ )I}M
Fig. 6. Responses of the controlled power system without compensation ηi,k
K = diag{Kii }M , C = diag{Cii }M under deception attacks.
G = diag{L11 C11 , 0, . . . , 0}
⎡ ⎤
A1 0 −B
Ā = ⎣ G A1 − G 0 ⎦ presented secure control strategies based on passive defense
θ KC − θ KC  or active compensation perform well and preserve satisfac-
tory stable behavior of the controlled power system. However,
where the decision parameters include a positive definite in the case of deception attacks, it is observed from Fig. 6
matrix P, controller gain matrices Kii , and a positive scalar θ . that the system performance without attack compensation is
It should be pointed out that the above centralized secure seriously degraded, and the degraded performance cannot be
control results can be further extended to decentralized and easily recovered even if attacks disappear. This further con-
distributed cases for different application scenarios. We next firms that active compensation plays a vital role in maintaining
validate the control performance of the above secure control system stability and safety.
schemes via a four area power system, as briefly demon-
strated in Fig. 3. The system parameters are the same with
ones in [11] except that the slopes of the power angle curves VI. C ONCLUSION AND C HALLENGING I SSUES
are P12 = P21 = 2 and P23 = P32 = P34 = P43 = 1, An overview of secure state estimation and control has
and the measurement matrix is selected as an identity one been carried out for CPSs of differential dynamical mod-
for the convenience of test which indicates a state feedback els. The latest development of secure state estimation has
controller. Furthermore, the deception attacks occur from 4.5 s been first examined based on different performance indica-
(i.e., k = 45) to 6.0 s (i.e., k = 60) and from 9.0 s (i.e., k = 90) tors and defense strategies. Then the recent secure control
to 10 s (i.e., k = 100) on Area 2 with a fixed amplitude 15 results have been reviewed and classified. Furthermore, two
and Area 3 with a fixed amplitude 12, while the occurring representative applications of secure estimation and control
instants of DoS attacks are the same. The test results are shown approaches for real-world CPSs, namely, water distribution
in Figs. 4–6. It can been seen from Figs. 4 and 5 that the systems and wide-area power systems, have been exemplified

Authorized licensed use limited to: University of Canberra. Downloaded on May 20,2021 at 11:45:52 UTC from IEEE Xplore. Restrictions apply.
DING et al.: SECURE STATE ESTIMATION AND CONTROL OF CPSs: A SURVEY 187

to enable an elementary analysis framework for securing mod- security-related issues. From a control-oriented perspec-
ern infrastructure. Undoubtedly, the existing theoretical results tive, this new line of research remains in its infancy and
are required to be reinspected in practical applications due deserves considerable attention.
mainly to various limitations from mathematical complexity, 5) Artificial Intelligence (AI)-Based Secure State
the assumptions of system and attack models, the reliability Estimation and Control Approaches: When CPSs
and the effectiveness of defense or compensation strategies, encounter sophisticated cyberattacks, the developed
and the integration requirement of attack detection and prede- secure estimation and detection algorithms should
termined system function. This article is by no means complete possess certain intellectualization to empower satisfied
to cover all research in this field. We highlight the following identification capabilities of unknown attack models
challenging issues that are worthy of further research. and be further integrated into the control design to
1) Efficient Location and Isolation Mechanisms of Attacks compensate for attack-incurred impacts. For example,
Across Cyber and Physical Domains: Physical compo- neural networks are designed in [27] as an intelligent
nents (e.g., sensors and actuators) and cyber components estimator for attack estimation and some variable
(e.g., computing and communication devices) of practi- structure controller is designed to compensate for the
cal CPSs are usually deployed in a spatially distributed effects of attacks. Albeit some recent progress along
way. Any corrupted data packet can be epidemically this line of research, how to design effective AI-
propagated over some communication topology and based countermeasures for CPSs against sophisticated
affect the dynamical behavior of the whole physical cyberattacks demands deep investigations.
system. As such, a challenge is to accurately locate
the attack sources and isolate the attack, allowing fur-
ther system performance analysis from the perspective R EFERENCES
of control engineering. [1] W. Abbas, A. Laszka, and X. Koutsoukos, “Improving network connec-
2) Complicated Time Series Analysis Subject to tivity and robustness using trusted nodes with application to resilient
consensus,” IEEE Control Netw. Syst., vol. 5, no. 4, pp. 2036–2048,
Simultaneous Cyberattacks and Communication Dec. 2018.
Scheduling and Network-Induced Phenomena: The [2] A. Alessandri and L. Zaccarian, “Stubborn state observers for linear
limited cyber resources of practical CPSs are required time-invariant systems,” Automatica, vol. 88, pp. 1–9, Feb. 2018.
to be properly scheduled to maximize utilization effi- [3] S. Amin, X. Litrico, S. Sastry, and A. Bayen, “Cyber security of water
SCADA systems—Part I: Analysis and experimentation of stealthy
ciency. This, however, will result in the ever-enhancing deception attacks,” IEEE Trans. Control Syst. Technol., vol. 21, no. 5,
complexity of attack-incurred impact analysis because pp. 1963–1970, Sep. 2013.
of complicated time series of cyberattacks and com- [4] S. Amin, X. Litrico, S. Sastry, and A. Bayen, “Cyber security of water
SCADA systems—Part II: Attack detection using enhanced hydrody-
munication scheduling mechanisms as well as various namic models,” IEEE Trans. Control Syst. Technol., vol. 21, no. 5,
network-induced phenomena, including delays/data pp. 1679–1693, Sep. 2013.
losses/packet disorders. The existing attack impact [5] L. An and G.-H. Yang, “LQ secure control for cyber-physical systems
against sparse sensor and actuator attacks,” IEEE Control Netw. Syst.,
analysis approaches based on switched system theory vol. 6, no. 2, pp. 833–841, Jun. 2019.
might be undesirable in this case due mainly to the [6] L. An and G.-H. Yang, “Secure state estimation against sparse sen-
randomness and uncertainty of cyberattacks. In the sor attacks with adaptive switching mechanism,” IEEE Trans. Autom.
Control, vol. 63, no. 8, pp. 2596–2603, Aug. 2018.
context of active defense, both the sparsity of data [7] O. Anubi and C. Konstantinou, “Enhanced resilient state estima-
and the uncertainty of transmission intervals seriously tion using data-driven auxiliary models,” IEEE Trans. Ind. Informat.,
obstruct the applicability of developed approaches, vol. 16, no. 1, pp. 639–647, Jan. 2020.
which demands careful consideration. [8] A. Ashok, M. Govindarasu, and J. Wang, “Cyber-physical attack-
resilient wide-area monitoring, protection, and control for the power
3) Scalable Secure State Estimation and Control Solutions grid,” Proc. IEEE, vol. 105, no. 7, pp. 1389–1407, Jul. 2017.
to Enable to Plug-and-Play CPSs: The scale or struc- [9] G. Casadei, D. Astolfi, A. Alessandri, and L. Zaccarian
ture of a practical CPS could suffer from connection “Synchronization in networks of identical nonlinear systems via
dynamic dead zones,” IEEE Control Syst. Lett., vol. 3, no. 3,
changes due to plug-and-play components. In this sense, pp. 667–672, Jul. 2019.
a centralized analysis and design approach necessitates [10] W. Chen, D. Ding, X. Ge, Q.-L. Han, and G. Wei, “H∞ containment
the global information of the CPS, which may severely control of multiagent systems under event-triggered communication
scheduling: The finite-horizon case,” IEEE Trans. Cybern., vol. 50,
restrict its applicability. As such, it is significant to no. 4, pp. 1372–1382, Apr. 2020.
develop novel scalable secure analysis and design strate- [11] W. Chen, D. Ding, H. Dong, and G. Wei, “Distributed resilient filtering
gies for the distributed implementation of large-scale for power systems subject to denial-of-service attacks,” IEEE Trans.
Syst., Man, Cybern., Syst., vol. 49, no. 8, pp. 1688–1697, Aug. 2019.
CPSs. [12] Y. Chen, S. Kar, and J. M. F. Moura, “Resilient distributed estima-
4) Data-Driven Secure State Estimation and Control tion: Sensor attacks,” IEEE Trans. Autom. Control, vol. 64, no. 9,
Technologies: The permeation of cyber resources in pp. 3772–3779, Sep. 2019.
CPSs poses inherent challenges to model physical [13] W.-L. Chin, C.-H. Lee, and T. Jiang, “Blind false data attacks against
AC state estimation based on geometric approach in smart grid com-
systems, which further makes existing model-based munications,” IEEE Trans. Smart Grid, vol. 9, no. 6, pp. 6298–6306,
security analysis and design approaches impracticable or Nov. 2018.
even invalid. Alternatively, data-driven state estimation [14] S. M. Dibaji, M. Pirani, D. B. Flamholz, A. M. Annaswamy,
K. H. Johansson, and A. Chakrabortty, “A systems and control per-
and control approaches [7], enabled by the increasing spective of CPS security,” Annu. Rev. Control, vol. 47, pp. 394–411,
use of IoT in CPSs, offer a great potential for addressing May 2019.

Authorized licensed use limited to: University of Canberra. Downloaded on May 20,2021 at 11:45:52 UTC from IEEE Xplore. Restrictions apply.
188 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO. 1, JANUARY 2021

[15] D. Ding, Q.-L. Han, Z. Wang, and X. Ge, “Recursive filtering [37] X. Ge, Q.-L. Han, L. Ding, Y.-L. Wang, and X.-M. Zhang, “Dynamic
of distributed cyber-physical systems with attack detection,” IEEE event-triggered distributed coordination control and its applications: A
Trans. Syst., Man, Cybern., Syst., early access, Jan. 10, 2019, survey of trends and techniques,” IEEE Trans. Syst., Man, Cybern.,
doi: 10.1109/TSMC.2019.2960541. Syst., vol. 50, no. 9, pp. 3112–3125, Sep. 2020.
[16] D. Ding, Z. Wang, D. W. C. Ho, and G. Wei, “Distributed recursive fil- [38] X. Ge, Q.-L. Han, X.-M. Zhang, L. Ding, and F. Yang, “Distributed
tering for stochastic systems under uniform quantizations and deception event-triggered estimation over sensor networks: A survey,” IEEE
attacks through sensor networks,” Automatica, vol. 78, pp. 231–240, Trans. Cybern., vol. 50, no. 3, pp. 1306–1320, Mar. 2020.
Apr. 2017. [39] W. He, Z. Mo, Q.-L. Han, and F. Qian, “Secure impulsive synchroniza-
[17] D. Ding, Z. Wang, and Q.-L. Han, “A set-membership approach to tion in Lipschitz-type multi-agent systems subject to deception attacks,”
event-triggered filtering for general nonlinear systems over sensor IEEE CAA J. Autom. Sinica, vol. 7, no. 5, pp. 1326–1334, Sep. 2020.
networks,” IEEE Trans. Autom. Control, vol. 65, no. 4, pp. 1792–1799, [40] W. He, F. Qian, Q.-L. Han, and G. Chen, “Almost sure stability of non-
Apr. 2020. linear systems under random and impulsive sequential attacks,” IEEE
[18] D. Ding, Z. Wang, Q.-L. Han, and G. Wei, “Neural-network-based Trans. Autom. Control, vol. 65, no. 9, pp. 3879–3886, Sep. 2020.
output-feedback control under Round-Robin scheduling protocols,” [41] N. Hou, Z. Wang, D. W. C. Ho, and H. Dong, “Robust partial-nodes-
IEEE Trans. Cybern., vol. 49, no. 6, pp. 2372–2384, Jun. 2019. based state estimation for complex networks under deception attacks,”
[19] D. Ding, Z. Wang, D. W. C. Ho, and G. Wei, “Observer-based event- IEEE Trans. Cybern., vol. 50, no. 6, pp. 2793–2802, Jun. 2020.
triggering consensus control for multiagent systems with lossy sensors [42] L. Hu, Z. Wang, Q.-L. Han, and X. Liu, “State estimation under
and cyber-attacks,” IEEE Trans. Cybern., vol. 47, no. 8, pp. 1936–1947, false data injection attacks: Security analysis and system protection,”
Aug. 2017. Automatica, vol. 87, pp. 176–183, Jan. 2018.
[20] D. Ding, Q.-L. Han, Y. Xiang, X. Ge, and X.-M. Zhang, “A survey [43] S. Hu, D. Yue, X. Xie, X. Chen, and X. Yin, “Resilient event-
on security control and attack detection for industrial cyber-physical triggered controller synthesis of networked control systems under
systems,” Neurocomputing, vol. 275, pp. 1674–1683, Jan. 2018. periodic DoS jamming attacks,” IEEE Trans. Cybern., vol. 49, no. 12,
[21] D. Ding, Z. Wang, Q.-L. Han, and G. Wei, “Security control for pp. 4271–4281, Dec. 2019.
discrete-time stochastic nonlinear systems subject to deception attacks,” [44] S. Hu, D. Yue, Q.-L. Han, X. Xie, X. Chen, and C. Dou, “Observer-
IEEE Trans. Syst., Man, Cybern., Syst., vol. 48, no. 5, pp. 779–789, based event-triggered control for networked linear systems subject
May 2018. to denial-of-service attacks,” IEEE Trans. Cybern., vol. 50, no. 5,
[22] L. Ding, Q.-L. Han, and X.-M. Zhang, “Distributed secondary con- pp. 1952–1964, May 2020.
trol for active power sharing and frequency regulation in islanded [45] S. Knorn and A. Teixeira, “Effects of jamming attacks on a control
microgrids using an event-triggered communication mechanism,” IEEE system with energy harvesting,” IEEE Control Syst. Lett., vol. 3, no. 4,
Trans. Ind. Informat., vol. 15, no. 7, pp. 3910–3922, Jul. 2019. pp. 829–834, Oct. 2019.
[23] L. Ding, Q.-L. Han, B. Ning, and D. Yue, “Distributed resilient [46] J. Li, Y. Liu, T. Chen, Z. Xiao, Z. Li, and J. Wang, “Adversarial attacks
finite-time secondary control for heterogeneous battery energy storage and defenses on cyber–physical systems: A survey,” IEEE Internet
systems under denial-of-service attacks,” IEEE Trans. Ind. Informat., Things J., vol. 7, no. 6, pp. 5103–5115, Jun. 2020.
vol. 16, no. 7, pp. 4909–4919, Jul. 2020. [47] C. S. Liu, J. Wu, C. N. Long, and Y. B. Wang, “Dynamic state recov-
[24] V. Do, L. Fillatre, I. Nikiforov, and P. Willett, “Security of SCADA ery for cyber-physical systems under switching location attacks,” IEEE
systems against cyber–physical attacks,” IEEE Aerosp. Electron. Syst. Control Netw. Syst., vol. 4, no. 1, pp. 14–22, Mar. 2017.
Mag., vol. 32, no. 5, pp. 28–45, May 2017. [48] J. Liu, Y. Gu, L. Zha, Y. Liu, and J. Cao, “Event-triggered H∞
[25] V. Do, “Sequential detection and isolation of cyber-physical attacks on load frequency control for multiarea power systems under hybrid
SCADA systems,” Ph.D. Dissertation, Univ. Technol., Troyes, Troyes, cyber attacks,” IEEE Trans. Syst., Man, Cybern., Syst., vol. 49, no. 8,
France, 2015. pp. 1665–1678, Aug. 2019.
[26] D. Du, C. Zhang, H. Wang, X. Li, H. Hu, and T. Yang, “Stability analy- [49] J. Liu, X. Lu, and J. Wang, “Resilience analysis of DC microgrids
sis of token-based wireless networked control systems under deception under denial of service threats,” IEEE Trans. Power Syst., vol. 34,
attacks,” Inf. Sci., vol. 459, pp. 168–182, Aug. 2018. no. 4, pp. 3199–3208, Jul. 2019.
[27] F. Farivar, M. Haghighi, A. Jolfaei, and M. Alazab, “Artificial intelli- [50] K. Liu, H. Guo, Q. Zhang, and Y. Xia, “Distributed secure filtering
gence for detection, estimation, and compensation of malicious attacks for discrete-time systems under Round-Robin protocol and decep-
in nonlinear cyber-physical systems and industrial IoT,” IEEE Trans. tion attacks,” IEEE Trans. Cybern., vol. 50, no. 8, pp. 3571–3580,
Ind. Informat., vol. 16, no. 4, pp. 2716–2725, Apr. 2020. Aug. 2020.
[28] H. Fawzi, P. Tabuada, and S. Diggavi, “Secure estimation and con- [51] S. Liu, Z. Hu, X. Wang, and L. Wu, “Stochastic stability analysis
trol for cyber-physical systems under adversarial attacks,” IEEE Trans. and control of secondary frequency regulation for islanded microgrids
Autom. Control, vol. 59, no. 6, pp. 1454–1467, Jun. 2014. under random denial of service attacks,” IEEE Trans. Ind. Informat.,
[29] S. Feng and P. Tesi, “Resilient control under denial-of-service: Robust vol. 15, no. 7, pp. 4066–4075, Jul. 2019.
design,” Automatica, vol. 79, pp. 42–51, May 2017. [52] A.-Y. Lu and G.-H. Yang, “Switched projected gradient descent
[30] Z. Feng and G. Hu, “Secure cooperative event-triggered control of algorithms for secure state estimation under sparse sensor attacks,”
linear multiagent systems under DoS attacks,” IEEE Trans. Control Automatica, vol. 103, pp. 503–514, May 2019.
Syst. Technol., vol. 28, no. 3, pp. 741–752, May 2020. [53] A.-Y. Lu and G.-H. Yang, “Secure switched observers for cyber-
[31] G. Franzè, F. Tedesco, and W. Lucia, “Resilient control for cyber- physical systems under sparse sensor attacks: A set cover approach,”
physical systems subject to replay attacks,” IEEE Control Syst. Lett., IEEE Trans. Autom. Control, vol. 64, no. 9, pp. 3949–3955, Sep. 2019.
vol. 3, no. 4, pp. 984–989, Oct. 2019. [54] A.-Y. Lu and G.-H. Yang, “Observer-based control for cyber-physical
[32] W. Fu, J. Qin, Y. Shi, W. Zheng, and Y. Kang, “Resilient consen- systems under denial-of-service with a decentralized event-triggered
sus of discrete-time complex cyber-physical networks under deception scheme,” IEEE Trans. Cybern., vol. 50, no. 12, pp. 4886–4895,
attacks,” IEEE Trans. Ind. Informat., vol. 16, no. 7, pp. 4868–4877, Dec. 2020, doi: 10.1109/TCYB.2019.2944956.
Jul. 2020. [55] L. Ma, Z. Wang, Q.-L. Han, and H.-K. Lam, “Variance-constrained
[33] Y. Gao, G. Sun, J. Liu, Y. Shi, and L. Wu, “State estimation and self- distributed filtering for time-varying systems with multiplicative noises
triggered control of CPSs against joint sensor and actuator attacks,” and deception attacks over sensor networks,” IEEE Sensors J., vol. 17,
Automatica, vol. 113, Mar. 2020, Art. no. 108687. no. 7, pp. 2279–2288, Apr. 2017.
[34] X. Ge, Q.-L. Han, and Z. Wang, “A dynamic event-triggered transmis- [56] R. Ma, P. Shi, Z. Wang, and L. Wu, “Resilient filtering for cyber-
sion scheme for distributed set-membership estimation over wireless physical systems under denial-of-service attacks,” Int. J. Robust
sensor networks,” IEEE Trans. Cybern., vol. 49, no. 1, pp. 171–183, Nonlinear Control, vol. 30, pp. 1754–1769, Mar. 2020.
Jan. 2019. [57] F. Miao, Q. Zhu, M. Pajic, and G. J. Pappas, “A hybrid stochastic game
[35] X. Ge, Q.-L. Han, M. Zhong, and X.-M. Zhang, “Distributed Krein for secure control of cyber-physical systems,” Automatica, vol. 93,
space-based attack detection over sensor networks under deception pp. 55–63, Jul. 2018.
attacks,” Automatica, vol. 109, Nov. 2019, Art. no. 108557. [58] S. Mishra, Y. Shoukry, N. Karamchandani, S. N. Diggavi, and
[36] X. Ge, Q.-L. Han, X.-M. Zhang, D. Ding, and F. Yang, “Resilient and P. Tabuada, “Secure state estimation against sensor attacks in the pres-
secure remote monitoring for a class of cyber-physical systems against ence of noise,” IEEE Control Netw. Syst., vol. 4, no. 1, pp. 49–59,
attacks,” Inf. Sci., vol. 512, pp. 1592–1605, Feb. 2020. Mar. 2017.

Authorized licensed use limited to: University of Canberra. Downloaded on May 20,2021 at 11:45:52 UTC from IEEE Xplore. Restrictions apply.
DING et al.: SECURE STATE ESTIMATION AND CONTROL OF CPSs: A SURVEY 189

[59] Y. Mo, R. Chabukswar, and B. Sinopoli, “Detecting integrity attacks on [80] S. Sundaram and C. Hadjicostis, “Distributed function calculation via
SCADA systems,” IEEE Trans. Control Syst. Technol., vol. 22, no. 4, linear iterative strategies in the presence of malicious agents,” IEEE
pp. 1396–1407, Jul. 2014. Trans. Autom. Control, vol. 56, no. 7, pp. 1495–1508, Jul. 2011.
[60] Y. Mo, S. Weerakkody, and B. Sinopoli, “Physical authentication [81] H. Song, D. Ding, H. Dong, and Q.-L. Han, “Distributed maximum
of control systems: Designing watermarked control inputs to detect correntropy filtering for stochastic nonlinear systems under decep-
counterfeit sensor outputs,” IEEE Control Syst. Mag., vol. 35, no. 1, tion attacks,” IEEE Trans. Cybern., early access, Sep. 16, 2020,
pp. 93–109, Feb. 2015. doi: 10.1109/TCYB.2020.3016093.
[61] H. Modares, B. Kiumarsi, F. L. Lewis, F. Ferrese, and A. Davoudi, [82] A. Teixeira, K. C. Sou, H. Sandberg, and K. H. Johansson, “Secure con-
“Resilient and robust synchronization of multiagent systems under trol systems: A quantitative risk management approach,” IEEE Contr.
attacks on sensors and actuators,” IEEE Trans. Cybern., vol. 50, no. 3, Syst. Mag., vol. 35, no. 1, pp. 24–45, Feb. 2015.
pp. 1240–1250, Mar. 2020. [83] A. Teixeira, I. Shames, H. Sandgerg, and K. Johansson, “A secure con-
[62] R. Moghadam and H. Modares, “Resilient autonomous control of dis- trol framework for resource-limited adversaries,” Automatica, vol. 51,
tributed multiagent systems in contested environments,” IEEE Trans. pp. 135–148, Jan. 2015.
Cybern., vol. 49, no. 11, pp. 3957–3967, Nov. 2019. [84] K. Tsang, M. Huang, K. Johansson, and L. Shi, “Sparse linear injection
[63] E. Mousavinejad, F. Yang, Q.-L. Han, X. Ge, and L. Vlacic, attack on multi-agent consensus control systems,” IEEE Control Syst.
“Distributed cyber attacks detection and recovery mechanism for vehi- Lett., vol. 5, no. 2, pp. 665–670, Apr. 2021.
cle platooning,” IEEE Trans. Intell. Transp. Syst., vol. 21, no. 9, [85] M. Wolf and D. Serpanos, “Safety and security in cyber-physical
pp. 3821–3834, Sep. 2020. systems and Internet-of-Things systems,” Proc. IEEE, vol. 106, no. 1,
[64] A. S. Musleh, G. Chen, and Z. Dong, “A survey on the detection pp. 9–20, Jan. 2018.
algorithms for false data injection attacks in smart grids,” IEEE Trans. [86] S. Xiao, Q.-L. Han, X. Ge, and Y. Zhang, “Secure distributed finite-time
Smart Grid, vol. 11, no. 3, pp. 2218–2234, May 2020. filtering for positive systems over sensor networks under decep-
[65] Y. Nakahira and Y. Mo, “Attack-resilient H2 , H∞ , and 1 state esti- tion attacks,” IEEE Trans. Cybern., vol. 50, no. 3, pp. 1220–1229,
mator,” IEEE Trans. Autom. Control, vol. 63, no. 12, pp. 4353–4360, Mar. 2020.
Dec. 2018.
[87] S. Xiao, Q.-L. Han, X. Ge, and Y. Zhang, “Distributed
[66] Y. Ni, Z. Guo, Y. Mo, and L. Shi, “On the performance analysis of resilient estimator design for positive systems under topologi-
reset attack in cyber-physical systems,” IEEE Trans. Autom. Control, cal attacks,” IEEE Trans. Cybern., early access, Apr. 17, 2020,
vol. 65, no. 1, pp. 419–425, Jan. 2020. doi: 10.1109/TCYB.2020.2981646.
[67] M. Pajic, J. Weimer, N. Bezzo, O. Sokolsky, G. J. Pappas, and I. Lee, [88] C.-H. Xie and G.-H. Yang, “Secure estimation for cyber-physical
“Design and implementation of attack-resilient cyberphysical systems: systems with adversarial attacks and unknown inputs: An L2 -
With a focus on attack-resilient state estimators,” IEEE Control Syst. gain method,” Int. J. Robust Nonlinear Control, vol. 28, no. 4,
Mag., vol. 37, no. 2, pp. 66–81, Apr. 2017. pp. 2131–2143, Apr. 2018.
[68] M. Pajic, I. Lee, and G. J. Pappas, “Attack-resilient state estimation
[89] X. Xie, Z. Yang, and X. Mu, “Observer-based consensus control of non-
for noisy dynamical systems,” IEEE Control Netw. Syst., vol. 4, no. 1,
linear multiagent systems under semi-Markovian switching topologies
pp. 82–92, Mar. 2017.
and cyber attacks,” Int. J. Robust Nonlinear Control, vol. 30, no. 14,
[69] G. Park, C. Lee, H. Shim, Y. Eun, and K. H. Johansson, “Stealthy pp. 5510–5528, Sep. 2020.
adversaries against uncertain cyber-physical systems: Threat of robust
[90] W. Xu, G. Hu, D. W. C. Ho, and Z. Feng, “Distributed secure coopera-
zero-dynamics attack,” IEEE Trans. Autom. Control, vol. 64, no. 12,
tive control under denial-of-service attacks from multiple adversaries,”
pp. 4907–4919, Dec. 2019.
IEEE Trans. Cybern., vol. 50, no. 8, pp. 3458–3467, Aug. 2020.
[70] F. Pasqualetti, A. Bicchi, and F. Bullo, “Consensus computation in
[91] H. Yang, H. Xu, Y. Xia, and J. Zhang, “Stability analysis on networked
unreliable networks: A system theoretic approach,” IEEE Trans. Autom.
control systems under double attacks with predictive control,” Int. J.
Control, vol. 57, no. 1, pp. 90–104, Jan. 2012.
Robust Nonlinear Control, vol. 30, no. 4, pp. 1549–1563, Mar. 2020.
[71] C. Peng, J. Li, and M. Fei, “Resilient event-triggering H∞ load
frequency control for multi-area power systems with energy-limited [92] W. Yang, Y. Zhang, G. Chen, C. Yang, and L. Shi, “Distributed filtering
DoS attacks,” IEEE Trans. Power Syst., vol. 32, no. 5, pp. 4110–4118, under false data injection attacks,” Automatica, vol. 102, pp. 34–44,
Sep. 2017. Apr. 2019.
[72] C. Peng, H. Sun, M. Yang, and Y.-L. Wang, “A survey on security com- [93] H. Yuan, Y. Xia, and H. Yang, “Resilient state estimation of cyber-
munication and control for smart grids under malicious cyber attacks,” physical system with multichannel transmission under DoS attack,”
IEEE Trans. Syst., Man, Cybern., Syst., vol. 49, no. 8, pp. 1554–1569, IEEE Trans. Syst., Man, Cybern., Syst., early access, Jan. 21, 020,
Aug. 2019. doi: 10.1109/TSMC.2020.2964586.
[73] C. Persis and P. Tesi, “Input-to-state stabilizing control under denial-of- [94] Y. Yuan, H. Yuan, L. Guo, H. Yang, and S. Sun, “Resilient con-
service,” IEEE Trans. Autom. Control, vol. 60, no. 11, pp. 2930–2944, trol of networked control system under DoS attacks: A unified game
Nov. 2015. approach,” IEEE Trans. Ind. Informat., vol. 12, no. 5, pp. 1786–1794,
[74] P. S. P. Pessim and M. J. Lacerda, “State-feedback control for cyber- Oct. 2016.
physical LPV systems under DoS attacks,” IEEE Control Syst. Lett., [95] Y. Yuan, H. Yuan, D. W. C. Ho, and L. Guo, “Resilient control of
vol. 5, no. 3, pp. 1043–1048, Jul. 2021. wireless networked control system under denial-of-service attacks: A
[75] B. Shen, Z. Wang, D. Wang, and Q. Li, “State-saturated recur- cross-layer design approach,” IEEE Trans. Cybern., vol. 50, no. 1,
sive filter design for stochastic time-varying nonlinear com- pp. 48–60, Jan. 2020.
plex networks under deception attacks,” IEEE Trans. Neural [96] J. Zhang, J. Sun, and C. Zhang, “Stochastic game in linear quadratic
Netw. Learn. Syst, vol. 31, no. 10, pp. 3788–3800, Oct. 2020, Gaussian control for wireless networked control systems under DoS
doi: 10.1109/TNNLS.2019.2946290. attacks,” IEEE Trans. Syst., Man, Cybern., Syst., early access, Aug. 5,
[76] X.-Y. Shen and X.-J. Li, “Data-driven output-feedback LQ secure 2020, doi: 10.1109/TSMC.2020.3010515.
control for unknown cyber-physical systems against sparse actuator [97] R. Zhang and P. Venkitasubramaniam, “False data injection and detec-
attacks,” IEEE Trans. Syst., Man, Cybern., Syst., early access, Dec. 20, tion in LQG systems: A game theoretic approach,” IEEE Control Netw.
2019, doi: 10.1109/TSMC.2019.2957146. Syst., vol. 7, no. 1, pp. 338–348, Mar. 2020.
[77] Y. Shoukry, P. Nuzzo, A. Puggelli, A. L. Sangiovanni-Vincentelli, [98] X.-M. Zhang, Q.-L. Han, X. Ge, and D. Ding, “Networked control
S. A. Seshia, and P. Tabuada, “Secure state estimation for cyber- systems: A survey of trends and techniques,” IEEE CAA J. Autom.
physical systems under sensor attacks: A satisfiability modulo the- Sinica, vol. 7, no. 1, pp. 1–17, Jan. 2020.
ory approach,” IEEE Trans. Autom. Control, vol. 62, no. 10, [99] X.-M. Zhang, Q.-L. Han, X. Ge, and L. Ding, “Resilient control design
pp. 4917–4932, Oct. 2017. based on a sampled-data model for a class of networked control systems
[78] M. Showkatbakhsh, Y. Shoukry, S. N. Diggavi, and P. Tabuada, under denial-of-service attacks,” IEEE Trans. Cybern., vol. 50, no. 8,
“Securing state reconstruction under sensor and actuator attacks: pp. 3616–3626, Aug. 2020.
Theory and design,” Automatica, vol. 116, Jun. 2020, Art. no. 108920. [100] D. Zhao, Z. Wang, D. W. C. Ho, and G. Wei, “Observer-based PID
[79] Y. Sun, D. Ding, H. Dong, and H. Liu, “Event-based resilient filtering security control for discrete time-delay systems under cyber-attacks,”
for stochastic nonlinear systems via innovation constraints,” Inf. Sci., IEEE Trans. Syst., Man, Cybern., Syst., early access, Dec. 5, 2019,
vol. 546, pp. 512–525, Feb. 2021. doi: 10.1109/TSMC.2019.2952539.

Authorized licensed use limited to: University of Canberra. Downloaded on May 20,2021 at 11:45:52 UTC from IEEE Xplore. Restrictions apply.
190 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO. 1, JANUARY 2021

[101] Q. Zhu and T. Basar, “Game-theoretic methods for robustness, security, Xiaohua Ge (Member, IEEE) received the B.Eng.
and resilience of cyberphysical control systems: Games-in-games prin- degree in electronics and information engineering
ciple for optimal cross-layer resilient control systems,” IEEE Control from Nanchang Hangkong University, Nanchang,
Syst. Mag., vol. 35, no. 1, pp. 46–65, Feb. 2015. China, in 2008, the M.Eng. degree in control
[102] Y. Zhu and W. Zheng, “Observer-based control for cyber-physical theory and control engineering from Hangzhou
systems with periodic DoS attacks via a cyclic switching strategy,” Dianzi University, Hangzhou, China, in 2011, and
IEEE Trans. Autom. Control, vol. 65, no. 8, pp. 3714–3721, Aug. 2020. the Ph.D. degree in computer engineering from
[103] L. Zou, Z. Wang, Q.-L. Han, and D. Zhou, “Recursive filtering for time- Central Queensland University, Rockhampton, QLD,
varying systems with random access protocol,” IEEE Trans. Autom. Australia, in 2014.
Control, vol. 64, no. 2, pp. 720–727, Feb. 2019. From 2011 to 2013, he was a Research Assistant
[104] L. Zou, Z. Wang, Q.-L. Han, and D. Zhou, “Moving horizon estimation with the Centre for Intelligent and Networked
for networked time-delay systems under Round-Robin protocol,” IEEE Systems, Central Queensland University, where he was a Research Fellow
Trans. Autom. Control, vol. 64, no. 12, pp. 5191–5198, Dec. 2019. in 2014. From 2015 to 2017, he was a Research Fellow with the Griffith
[105] S. Zuo, O. Beg, F. L. Lewis, and A. Davoudi, “Resilient networked AC School of Engineering, Griffith University, Gold Coast, QLD, Australia. He
microgrids under unbounded cyber attacks,” IEEE Trans. Smart Grid, is currently a Senior Lecturer with the School of Software and Electrical
vol. 11, no. 5, pp. 3785–3794, Sep. 2020. Engineering, Swinburne University of Technology, Melbourne, VIC, Australia.
[106] S. Zuo, T. Altun, F. L. Lewis, and A. Davoudi, “Distributed resilient His research interests include distributed estimation over sensor networks,
secondary control of DC microgrids against unbounded attacks,” IEEE distributed coordination in multiagent systems, and security and privacy
Trans. Smart Grid, vol. 11, no. 5, pp. 3850–3859, Sep. 2020. preserving in cyber-physical systems.
Dr. Ge was a Guest Editor of the IEEE T RANSACTION ON
S YSTEMS , M AN , AND C YBERNETICS : S YSTEMS, Information Sciences,
Derui Ding (Senior Member, IEEE) received Neurocomputing, and Kybernetika.
the B.Sc. degree in industry engineering and the
M.Sc. degree in detection technology and automa-
tion equipment from Anhui Polytechnic University,
Wuhu, China, in 2004 and 2007, respectively, and
the Ph.D. degree in control theory and control engi-
neering from Donghua University, Shanghai, China,
in 2014.
From July 2007 to December 2014, he was
a Teaching Assistant and then a Lecturer with
the Department of Mathematics, Anhui Polytechnic
University. He is currently a Senior Research Fellow with the School of
Software and Electrical Engineering, Swinburne University of Technology,
Melbourne, VIC, Australia. From June 2012 to September 2012, he was
a Research Assistant with the Department of Mechanical Engineering,
University of Hong Kong, Hong Kong. From March 2013 to March 2014,
he was a Visiting Scholar with the Department of Information Systems and
Computing, Brunel University London, Uxbridge, U.K. He has published
around 80 papers in refereed international journals. His research interests
include nonlinear stochastic control and filtering, as well as multiagent systems
and sensor networks.
Dr. Ding is serving as an Associate Editor for Neurocomputing and IET
Control Theory & Applications. He is also a very active reviewer for many
international journals.

Qing-Long Han (Fellow, IEEE) received the

B.Sc. degree in mathematics from Shandong
Normal University, Jinan, China, in 1983, and the
M.Sc. and Ph.D. degrees in control engineering
from the East China University of Science and
Technology, Shanghai, China, in 1992 and 1997,
respectively.
He is a Pro Vice-Chancellor (Research Quality)
and a Distinguished Professor with the Swinburne Jun Wang (Fellow, IEEE) received the B.S. degree
University of Technology, Melbourne, VIC, in electrical engineering and the M.S. degree in
Australia. He held various academic and man- systems engineering from the Dalian University
agement positions with Griffith University, Brisbane, QLD, Australia, and of Technology, Dalian, China, in 1982 and 1985,
Central Queensland University, Rockhampton, QLD, Australia. His research respectively, and the Ph.D. degree in systems engi-
interests include networked control systems, multiagent systems, time-delay neering from Case Western Reserve University,
systems, smart grids, unmanned surface vehicles, and neural networks. Cleveland, OH, USA, in 1991.
Prof. Han received the 2020 IEEE Systems, Man, and Cybernetics Society He is a Chair Professor with the Department of
Andrew P. Sage Best Transactions Paper Award, the 2020 IEEE Industrial Computer Science, City University of Hong Kong,
Electronics Society IEEE T RANSACTIONS ON I NDUSTRIAL I NFORMATICS Hong Kong. Prior to this position, he held various
Outstanding Paper Award, and the 2019 IEEE Systems, Man, and Cybernetics academic positions with the Dalian University of
Society Andrew P. Sage Best Transactions Paper Award. He is a Highly Technology; Case Western Reserve University; University of North Dakota,
Cited Researcher according to Clarivate Analytics. He is a Fellow of the Grand Forks, ND, USA; and the Chinese University of Hong Kong, Hong
Institution of Engineers Australia. He is one of Australia’s Top 5 Lifetime Kong.
Achievers (Research Superstars) in the discipline area of Engineering Prof. Wang recipient of the Outstanding Paper Award of IEEE
and Computer Science by the Australians 2020 Research Magazine. T RANSACTIONS ON N EURAL N ETWORKS, the Outstanding Achievement
He is a Co-Editor of Australian Journal of Electrical and Electronic Award from the Asia Pacific Neural Network Assembly, the Neural Networks
Engineering, an Associate Editor for 12 international journals, including Pioneer Award from the IEEE Computational Intelligence Society, and the
the IEEE T RANSACTIONS ON C YBERNETICS, IEEE T RANSACTIONS Norbert Wiener Award and Outstanding Contribution Award from the IEEE
ON I NDUSTRIAL I NFORMATICS , IEEE Industrial Electronics Magazine, Systems, Man and Cybernetics Society, among others. He was the Editor-in-
IEEE/CAA J OURNAL OF AUTOMATICA S INICA, Control Engineering Chief of the IEEE T RANSACTIONS ON C YBERNETICS. He is a Fellow of the
Practice, and Information Sciences, and a Guest Editor for 13 Special Issues. International Association for Pattern Recognition.

Authorized licensed use limited to: University of Canberra. Downloaded on May 20,2021 at 11:45:52 UTC from IEEE Xplore. Restrictions apply.