IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 9, NO.

6, JUNE 2022 1099

Letter

Attack-Resilient Control Against FDI [12] to ensure that CPSs are stochastic finite-time bounded when subj-
ected to random injection attacks. Note that the aforementioned
Attacks in Cyber-Physical Systems studies are under the same assumption, i.e., the system states can be
directly measured. Nevertheless, this strict requirement is normally
Bo Chen, Member, IEEE, Yawen Tan, Zhe Sun, and difficult to be satisfied in reality. Consequently, observer-based SMC
Li Yu, Member, IEEE has drawn more and more attention. For instance, an observer-based
event-triggered SMC method is developed for nonlinear networked
control systems subjected to cyber attacks [13]. In [14], a novel
Dear editor, technique based on neural network is presented to estimate the attack
signals for attenuating the negative effect, and an adaptive SMC
This letter is concerned with the control of cyber-physical systems method is designed to guarantee the system’s security.
(CPSs) in the presence of malicious false data injection (FDI) attacks Motivated by the aforementioned research, this letter studies the
on actuators. The FDI attacks on actuators may result in faults of control of CPSs under FDI attacks. An unknown input observer
actuators or even the instability of CPSs. To tackle this problem, an (UIO) is designed to jointly estimate the system states and attack
unknown input observer (UIO) is proposed to estimate the system signals, and a DSMC law is proposed to attenuate the impact of FDI
states and attack signals. For the aim of suppressing the impact of attacks and maintain the stability of the control system. Finally, the
FDI attacks, a discrete-time sliding mode control (DSMC) algorithm effectiveness of the proposed DSMC law is verified from two
is correspondingly put forward, where its reaching law is constructed simulation examples.
based on the n-th order difference of the estimation of attack signals. The main contributions of this letter are summarized as follows:
Finally, two simulation instances are presented to show the effec- 1) For a system with unknown states, an UIO is put forward such
tiveness of the proposed method. that the system states and FDI attack signals can be accurately
Introduction: Cyber-physical system (CPS) plays a more and estimated.
more pivotal role in the era of Industry 4.0 and attracts increasing 2) A DSMC algorithm is accordingly designed by introducing the
attention from academy and industry profiting by the advantages of high-order terms with respect to the estimation of attack signals,
networks such as low cost, less wiring and convenient maintenance which further reduces the width of the quasi-sliding mode domain
[1]–[3]. However, the introduction of communication networks breaks (QSMD).
the closeness of conventional physical systems, and thus results in Notations: Rn and Rn×m denote the n -dimensional and n × m
the threat of cyber attacks [4]–[6]. Among various cyber attacks, FDI dimensional Euclidean spaces, respectively; the superscript T
attack is a typical and threatening one, whose working principle is to represents the transpose operation; In denotes an n-dimensional
break the integrity and availability of data by injecting erroneous identity matrix.
signals [7]. Therefore, it is of great significance to research effective Problem formulation: Consider a linear discrete-time CPS where
defense strategies for CPSs. the actuator is subjected to FDI attacks. The system’s expression is
In order to maintain the security and improve the robustness of shown as follows:
{
CPSs under cyber attacks, the concept and methods of fault tolerant x(k + 1) = Ax (k) + B [u (k) + a (k)]
control (FTC) are proposed. According to different working princip- (1)
les, FTC is classified into two types, i.e., active ones and passive y (k) = Cx (k)
ones [8]. Active FTC utilizes attack-detection method to identify where x ∈ Rn x is the system state; u ∈ Rl denotes the control input;
attacks or faults, and then appropriately manipulate the control input y ∈ Rny is the measurement signal; a ∈ Rl represents the FDI attack
to compensate for the corresponding negative effect [9]. On the other signal; A, B and C are constant matrices of appropriate dimensions. T
hand, passive FTC normally ensures the stability of a control system and k denote the sampling period and time step, respectively. For this
by considering and addressing the worst case. In contrast with system, we assume that the pair (A, B) is controllable, and the pair
passive FTC, the active FTC input contains the estimating (A,C) is observable.
information of attacks or faults, which makes the control system less A robust sliding mode controller can be designed such that the
conservative. Hence, the investigation of active FTC has aroused system is capable of maintaining stable and effective after being
keen interest from researchers and engineers. attacked, where the corresponding sliding surface s is preliminarily
As a robust and effective nonlinear control strategy, sliding mode defined as
control (SMC) is recognized as one of the most competent tools in
s (k) = Gx (k) (2)
dealing with uncertain systems owing to its robustness or even
insensitivity to perturbations. Hence, SMC has been introduced for where G ∈ Rl×nx and GB is nonsingular.
the purposed of ensuring the security of CPSs [10]. Specifically, Li Problem 1: Design a suitable UIO and a DSMC law to ensure that
et al. proposed an adaptive SMC law with a discontinuous input term when the system is suffering from FDI attacks, the system
to defense actuator attacks [11]. An SMC method is put forward by trajectories can converge to the sliding surface, and subsequently
converge to the origin along the sliding surface.
Corresponding author: Zhe Sun. Control design: In this section, an UIO is designed to estimate
Citation: B. Chen, Y. W. Tan, Z. Sun, and L. Yu, “Attack-resilient control unknown FDI attack signals, and a sliding mode controller is
against FDI attacks in cyber-physical systems,” IEEE/CAA J. Autom. Sinica, correspondingly designed to guarantee the stability of the control
vol. 9, no. 6, pp. 1099–1102, Jun. 2022.
system.
The authors are with the Department of Automation, Zhejiang University Initially, the state x (k) can be decoupled into a known input x̄1 (k)
of Technology, Hangzhou 310023, and also with the Institute of Cyberspace
Security, Zhejiang University of Technology, Hangzhou 310023, China (e- and an unknown input x̄2 (k). For this aim, Φ = [M B] and Λ =
mail: [email protected]; [email protected]; [email protected]; [CB N] are construct, where N ∈ Rny ×(ny −l) and M ∈ Rnx ×(nx −l) are
[email protected]). matrices chosen to satisfy that Φ and Λ are nonsingular [15]. Hence,
Color versions of one or more of the figures in this paper are available we have Λ−1 = [Λ1 Λ2 ]T with Λ1 ∈ Rl×ny and Λ2 ∈ R(ny −l)×ny .
online at http://ieeexplore.ieee.org. Multiplying Φ−1 and Λ−1 on both sides of the state and the mea-
Digital Object Identifier 10.1109/JAS.2022.105641 surement equation in (1) yields
1100 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 9, NO. 6, JUNE 2022



 x̄(k + 1) = Ā x̄ (k) + B̄ [u( k) + a(k )] own signals will also be higher. It can be seen that the proposed

 (3) DSMC law’s working principle is based on the statistical charact-
y (k) = C̄ x̄ (k) eristics of unknown signals.
{ Up to now, we have designed a DSMC law for the CPS under
Λ1 y (k) = Λ1CM x̄1 (k) + x̄2 (k) unknown FDI attacks. However, the stability of the control system is
(4)
Λ2 y (k) = Λ2CM x̄1 (k) required to be proved. Specifically, in our case, we need to subst-
[ ] antiate that the switch function s (k) from any initial state can reach
Ā11 Ā12
where x = Φ x̄ = Φ[ x̄1 x̄2 ]T , Ā = Φ−1 AΦ = , B̄ = the QSMD and subsequently remain in the QSMD.
Ā21 Ā22 Theorem 1: For the discrete-time CPS as shown in (1) with the
Φ−1 B = [0 Il ]T , C̄ = CΦ = [CM CB]. Combining (3) and (4) following assumption:
yields
 ξ = max∇n a (k) ≤ λ (11)


 x̄1 (k + 1) = Ã x̄1 (k) + Ey (k)
 (5) the system trajectories will converge to an O (T n )
QSMD by the

ȳ (k) = C̃ x̄ (k) proposed DSMC law, where the expression of Ω∆ is shown as below:
1
where à = Ā11 − Ā12 Λ1CM , E = Ā12 Λ1, ȳ (k) = Λ2 y (k), and Ω∆ = {si (k) | |si (k)| < λ + ξ} . (12)
C̃ = Λ2CM . A Luenberger observer can be designed while the Proof: Two cases of si (k) > 0 and si (k) < 0 are discussed here,
system matrices (Ã, C̃) is observable, In addition, the observability of respectively.
this system has been examined and affirmed in [16]. According to the For the first case si (k) > 0, from (9) we can obtain
concept of UIO proposed in [17], a Luenberger observer is constr-
ucted to estimate x̄1 (k) with x̄ˆ1, which signifies that x̄ˆ1 → x̄1 as si (k + 1) = (1 − qT ) si (k) − λ + ∇n ai (k)
t → ∞. The observer is shown as follows: < si (k) − λ + ∇n ai (k) < si (k) . (13)



 x̄ˆ1 (k + 1) = (Ã − LC̃) x̄ˆ1 (k) + (LΛ2 + E)y (k) As shown in (13), when si (k) > 0, the sliding variable si is
(k)


 x̄ˆ (k + 1) = Λ y(k + 1) − Λ CM x̄ˆ (k + 1)
(6) monotonically decreasing. Hence, there exists a situation where
2 1 1 1 si (k∗ ) > 0, and si (k∗ + 1) < 0. It is of necessity to investigate
where L ∈ R(n x −l)×(ny −l). Then, the estimated state can be obtained as whether si (k∗ + 1) will exceed Ω∆ in this case, i.e., whether
x̂ (k) = Φ x̄ˆ (k) . Thus, by substituting (6) into (3), we achieve the si (k∗ + 1) − (−λ − ξ) > 0. The deduction is shown as follows:
( ) ( ) ( )
estimation of the FDI attack signal a (k) as si k∗ + 1 − (−λ − ξ) = (1 − qT ) si k∗ − λ + ∇n ai k∗ + λ + ξ
â (k) = Λ1 y (k + 1) − Λ1CM x̄ˆ1 (k + 1) − Ā21 x̄ˆ1 (k) ( ) ( )
= (1 − qT ) si k∗ + ∇n ai k∗ + ξ
− Ā22 Λ1 y (k) + Ā22 Λ1CM x̄ˆ1 (k) − u (k) . ( )
(7) > (1 − qT ) si k∗ > 0. (14)
Rearranging (7), we can get a simplified equation as It can be seen that when si (k) > 0, the sliding variable si (k) will
â (k) = K1 y (k + 1) + K2 x̄ˆ1 (k) + K3 y (k) + K4 u (k) (8) converge into Ω∆.
For the second case si (k) < 0, we can prove in the same way. ■
where K2 = Λ1CMLΛ2CM − Λ1CM Ā11 + Λ1CM Ā12 Λ1CM− Though it has been verified that the switch function s (k) from any
Ā21 + Ā22 Λ1CM , K3 = −Λ1CMLΛ2 − Λ1CM Ā12 Λ1 − Ā22 Λ1, initial state will finally converge into the QSMD, it is of great
K1 = Λ1, K4 = −Il. The estimated state x̂ (k) and the estimated FDI necessity to prove that the trajectories will stay in the QSMD and not
attack signal â (k) will be utilized in the compensating procedure of escape from it.
the subsequent SMC design. Theorem 2: For the system (1) with the assumption (11) and under
For the SMC design, the first step is to construct an appropriate the DSMC law (10), when the switch function s (k) enters QSMD
sliding surface, which has been accomplished and embodied in (2). Ω∆, it will always stay in it.
Not only a suitable sliding variable is required, but also a reaching Proof: When s (k) enters the QSMD, there exist two situations,
law is of great importance. Herein, a reaching law based on [18] is namely, 0 ≤ s (k) ≤ λ + ξ and −λ − ξ ≤ s (k) ≤ 0 . Hence, we
adopted in our case, which is with the following expression: discuss this issue in two cases.
s (k + 1) = (1 − qT ) s (k) − λsign [s (k)] + ∇n a (k) (9) For the first case 0 ≤ s (k) ≤ λ + ξ , it can be derived that

where ∇a (k) = CBâ (k); ∇n a (k)

= ∇n−1 a (k) − ∇n−1 f
(k − 1); si (k + 1) = (1 − qT ) si (k) − λ + ∇n ai (k)
[ (
sign[s( k )] ≜ signs1 k), . . . , signsl (k)]T ; q is chosen to satisfy ≥ −λ + ∇n ai (k)
q > 0 and 0 < 1 − qT < 1; λ > 0 and n > 0 are control parameters
to be selected. ≥ − (λ + ξ) (15)
According to [19], we have the following lemma for the FDI attack and
signal a (k): si (k + 1) = (1 − qT ) si (k) − λ + ∇n ai (k)
Lemma 1: a (k) = O (T ) , a (k) − a (k − 1) = O(T 2 ), and a (k) − ( )
2a (k − 1) + a (k − 2) = O(T 3 ), T is the sampling time. ≤ λ + ξ − λ − ∇n ai (k)
From Lemma 1 it can be further concluded that ∇a (k) = O (T ); ≤ λ + ξ. (16)
∇2 a (k) = O(T 2 ) ; ∇n a (k) = O (T n ). Therefore, the upper bound of
the disturbance term in (9) can be decreased by augmenting the order For the second case −λ − ξ ≤ si (k) ≤ 0, the proof can be comp-
of ∇a (k). Subsequent analysis will further prove that the width of leted by a similar procedure.
QSMD can get reduced. Consequently, we achieve the conclusion that when the sliding
Combining (1), (2) and the reaching law (9), a DSMC law is variable s (k) enters the QSMD Ω∆, it stays inside Ω∆ and is unable
obtained, and the corresponding control input u (k) can be expressed as to escape from it. ■
[ To this end, we have presented the selection of the control
u (k) = − (GB)−1 GAx (k) − (1 − qT ) s (k) parameters of the DSMC.
∑n−1 1) Selection of λ: The parameter λ determines the width of QSMD
] and the capacity of the system to resist attacks. Increasing λ enhances
+ λsign (s( k )) + ∇i a(k − 1) . (10)
i=1 the attack-resistance ability but at the cost of bringing more
Remark 1: As the order of ∇a (k) increases, the width of QSMD chattering. Considering this tradeoff, we set λ = 0.2 .
will be shorten. Nevertheless, the constraint requirement for unkn- 2) Selection of q: The parameter q determines the convergence rate
CHEN et al.: ATTACK-RESILIENT CONTROL AGAINST FDI ATTACKS IN CPSS 1101

of the system. A larger q leads to a faster convergence rate but at the 6

cost of augmenting the risk of instability. To achieve a balance, the s(k)
4
parameter q is set as q = 5 .

Sliding mode s(k)

3) Selection of n: The parameter n denotes the order of the attack 2
signal. From Lemma 1, it can be seen that a larger n leads to a 0
narrower QSMD. However, increasing n also means more strict −2 0.2

conditions to be satisfied. Taking this tradeoff into consideration, we −4

0

set n = 5 . −0.2
100 120 140 160 180 200

Simulation results: A fourth-order inverted pendulum system [20] −6

0 20 40 60 80 100 120 140 160 180 200
is adopted as the controlled object whose dynamic model is as (1), k (Step)
where the system matrix, input matrix and measurement matrix are
   
 1
0.0991 0.0136 0.0004 
  0.0091  Fig. 2. Sliding mode dynamics in Case 1.
 0
0.9818 0.2789 0.0136  
 0.1822 
set as A =  , B =  
−0.0023 1.1598 0.1053 
 0  0.0232  4

FDI attack
2 Real FDI

signal
UIO
−0.0474 3.2764 1.1598
0 0.4732 0
 
 1 0 0 0  −2
20 40 60 80 100 120 140 160 180 200
 
and C =  0 0 1 0 , respectively. The initial state and the k (Step)
  1

The system
0 0 0 1 x1 x2 x3 x4

states
sampling period are set as x (0) = [0.98 0 0.2 0]T and T = 0.1 s. 0
Two simulations are carried out, where various FDI attack signals are −1
adopted to test the effectiveness and robustness of the UIO-based 0 20 40 60 80 100 120 140 160 180 200
SMC law. In Case 1, the FDI attack signal a (k) is set as k (Step)



 a (k) = −2cos (kπT ) , k < 50


Fig. 3. Estimation of FDI attack signal and system states in Case 2.

 a (k) = 2, 50 ≤ k < 125 (17)


 a (k) = 5sin (kπT ) , 125 ≤ k < 200. 1.4
s(k)
Sliding mode s(k) 1.2
Another relatively complicated FDI attack signal a(k) in Case 2 is 1.0
designed as 0.8 0.2

 0.6 0


 a (k) = sin (kπT ), k < 50 0.4


 0.2
−0.2
100 120 140 160 180 200



 1− 1

 a (k) = sin (kπT ) + e
 [(k−50)T ]2 , 50 ≤ k < 125 0
 (18) −0.2


 1− 1
0 20 40 60 80 100 120 140 160 180 200


 a (k) = sin (kπT ) + e [(k−50)T ]2 k (Step)




 1− 1
−e [(k− 125)T ]2 , 125 ≤ k < 200. Fig. 4. Sliding mode dynamics in Case 2.
By implementing the proposed algorithm, the simulation results are
obtained and shown in Figs. 1–4. Specifically, the estimation perfor- attack signals, and a DSMC algorithm is accordingly developed to
mance of the proposed UIO is shown in the upper parts of Figs. 1 attenuate the impact of FDI attacks, where the stability of the control
and 3, which turns out to be satisfactory. Under different attacks, the system is mathematically substantiated in detail. The simulation
estimated FDI attack signals are able to track the actual ones results evidently demonstrate the effectiveness and superiority of the
precisely. As shown in Fig. 2 , when the attack signal changes, presented UIO-based DSMC method. Note that our study in this
namely, when k = 125, the sliding variable deviates from the QSMD letter is concerned with a linear CPS model. However, most of the
for a pretty short time but then immediately converges back to the CPSs in reality are nonlinear ones. Hence, our future work is to study
QSMD and stays in it. The width of the QSMD is approximately 0.2 . nonlinear CPSs and put forward more robust and effective control
From Fig. 4 we can see that the proposed DSMC law is able to strategies.
guarantee the convergence within the QSMD, where the width of the Acknowledgments: This work was supported in part by the
QSMD is approximately 0.17. The lower parts of Figs. 1 and 3 National Natural Science Funds of China (61973277, 62073292) and
reveal that the system trajectories under the action of the presented the Zhejiang Provincial Natural Science Foundation of China
method converge to zero in a pretty short time and maintain stable (LR20F030004, LY20F020030).
thereafter.
Conclusions: This letter investigates the control of CPSs under
FDI attacks. An UIO is designed to estimate the system states and References
[1] D. Ding, Q. Han, X. Ge, and J. Wang, “ Secure state estimation and
5
FDI attack

control of cyber-physical systems: A survey,” IEEE Trans. Systems,

signal

0 Man, and Cybernetics: Systems, vol. 51, no. 1, pp. 176–190, 2020.
Real FDI UIO
−5 [2] B. Chen, D. W. C. Ho, G. Hu, and L. Yu, “Delay-dependent distributed
20 40 60 80 100 120 140 160 180 200 Kalman fusion estimation with dimensionality reduction in cyber-
k (Step) physical systems,” IEEE Trans. Cybernetics, 2021. DOI: 10.1109/
2
The system

TCYB.2021.3119461.
states

0 [3] S. Lai, B. Chen, T. Li, and L. Yu, “Packet-based state feedback control
x1 x2 x3 x4
under DOS attacks in cyber-physical systems,” IEEE Trans. Circuits
−2
0 20 40 60 80 100 120 140 160 180 200 and Systems II: Express Briefs, vol. 66, no. 9, pp. 1421–1425, 2019.
k (Step) [4] B. Chen, D. W. C. Ho, G. Hu, and L. Yu, “Secure fusion estimation for
bandwidth-constrained cyber-physical systems under replay attacks,”
Fig. 1. Estimation of FDI attack signal and system states in Case 1. IEEE Trans. Cybernetics, vol. 48, no. 6, pp. 1862–1876, 2018.
1102 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 9, NO. 6, JUNE 2022

[5] X. Ge, Q. Han, M. Zhong, and X. Zhang, “ Distributed krein space- [13] L. Zhang and G. Guo, “Observer-based adaptive event-triggered sliding
based attack detection over sensor networks under deception attacks,” mode control of saturated nonlinear networked systems with
Automatica, vol. 109, p. 108557, 2019. DOI: 10.1016/j.automatica.2019. cyberattacks,” Information Sciences, vol. 543, pp. 180–201, 2021.
108557.
[14] Z. Cao, Y. Niu, and Y. Zou, “Adaptive neural sliding mode control for
[6] B. Chen, G. Hu, D. W. Ho, and L. Yu, “ Distributed estimation and singular semi-Markovian jump systems against actuator attacks,” IEEE
control for discrete time-varying interconnected systems,” IEEE Trans. Trans. Systems, Man, and Cybernetics: Systems, vol. 51, no. 3,
Automatic Control, 2021. DOI: 10.1109/TAC.2021.3075198 pp. 1523–1533, 2019.
[7] L. Gao, B. Chen, and L. Yu, “ Fusion-based FDI attack detection in [15] M. Hou and P. Muller, “ Design of observers for linear systems with
cyberphysical systems,” IEEE Trans. Circuits and Systems II: Express unknown inputs,” IEEE Trans. Automatic Control , vol. 37, no. 6, pp.
Briefs, vol. 67, no. 8, pp. 1487–1491, 2019. 871–875, 1992.
[8] A. Abbaspour, S. Mokhtari, A. Sargolzaei, and K. K. Yen, “A survey on [16] M. R. Khalghani, J. Solanki, S. K. Solanki, M. H. Khooban, and A.
active fault-tolerant control systems,” Electronics, vol 9, no. 9, p. 1513,
Sargolzaei, “ Resilient frequency control design for microgrids under
2020. DOI: 10.3390/electronics9091513.
false data injection,” IEEE Trans. Industrial Electronics, vol. 68, no. 3,
[9] X. Yu and J. Jiang, “ A survey of fault-tolerant controllers based on pp. 2151–2162, 2020.
safety-related issues,” Annual Reviews in Control , vol. 39, pp. 46–57,
[17] J. Chen, R. J. Patton, and H. Zhang, “ Design of unknown input
2015.
observers and robust fault detection filters,” Int. J. Control , vol. 63,
[10] X. Huang, D, Zhai, and J. Dong, “ Adaptive integral sliding-mode no. 1, pp. 85–105, 1996.
control strategy of data-driven cyber-physical systems against a class of
actuator attacks,” IET Control Theory & Applications , vol. 12, no. 10, [18] W. Gao, Y. Wang, and A. Homaifa, “ Discrete-time variable structure
pp. 1440–1447, 2018. control systems,” IEEE Trans. Industrial Electronics , vol. 42, no. 2,
pp. 117–122, 1995.
[11] M. Li, Y. Chen, Y. Zhang, and Y. Liu, “Adaptive sliding-mode tracking
control of networked control systems with false data injection attacks,” [19] H. Ma, J. Wu, and Z. Xiong, “ A novel exponential reaching law of
Information Sciences, vol. 585, pp. 194–208, 2022. discrete-time sliding-mode control,” IEEE Trans. Industrial Electronics,
vol. 64, no. 5, pp. 3840–3850, 2017.
[12] Z. Cao, Y. Niu, and J. Song, “ Finite-time sliding-mode control of
Markovian jump cyber-physical systems against randomly occurring [20] H. Yang, Y. Xu, and J. Zhang, “ Event-driven control for networked
injection attacks,” IEEE Trans. Automatic Control, vol. 65, no. 3, control systems with quantization and Markov packet losses,” IEEE
pp. 1264–1271, 2019. Trans. Cybernetics, vol. 47, no. 8, pp. 2235–2243, 2017.