IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS, VOL.

7, 2021 465

Probability-Guaranteed Distributed Secure

Estimation for Nonlinear Systems Over Sensor
Networks Under Deception Attacks on Innovations
Lifeng Ma , Zidong Wang , Fellow, IEEE, Yun Chen , and Xiaojian Yi

Abstract—This paper studies the distributed secure state estima- such as military surveillance and environment detection [6],
tion problem for a class of general nonlinear systems over sensor [13], [29], [31], [32], [45], [46], [48]. A typical sensor network
networks under unknown deception attacks on innovations. At comprises a group of spatially dispersed sensing nodes with
each sensing node, an estimator is designed to generate the state
estimate by making use of the local measurements in combination basic capabilities of sensing, computing and communicating.
with the neighbours’ information shared via the communication To date, considerable research effort has been devoted to sensor
network. During the transmission of innovations among nodes, the networks from many different perspectives, and the distributed
data are maliciously falsified by adversaries in a random way. A state estimation problem has stirred particular interest, see
neural-network-based mechanism is put forward to approximate
e.g. [10]–[12], [19], [24], [34], [41] for some recent works.
the unknown falsified innovations with the aim to mitigate the
effects on the estimation performance. The objective of the ad- Compared with the traditional single-sensor systems, the sensor
dressed problem is to develop a distributed estimation approach to networks could collect more information in a comprehensive
jointly estimate the system states as well as the unknow deception and complementary way via the cooperation among individual
attacks, ensuring that the state estimation errors at each sensing nodes, thereby making the corresponding estimation algorithms
node reside within required ellipsoidal regions with a pre-specified
more robust, accurate and flexible. As is widely recognized,
probability. With the help of certain convex optimization methods,
we obtain sufficient conditions for the solvability of the addressed in the context of sensor networks, the scalability issue and
problem and the desired estimator gains can be iteratively com- the computational complexity are two major concerns for the
puted by solving a series of matrix inequalities. On basis of the conventional centralized estimation algorithms, and this gives
proposed framework, some optimization problems are presented rise to the necessity of developing the so-called distributed state
to determine sub-optimal estimator parameters from different
estimation/filtering problem where each individual sensing node
perspectives. Finally, the applicability of the developed algorithms
is validated via a numerical simulation example. provides the state estimates for the targets of interest through
making use of the local measurements in combination with the
Index Terms—Distributed estimation, set-membership state information shared by neighbours [4], [5], [15], [25], [26], [30],
estimation, deception attack, falsified innovations, probability-
guaranteed estimation, neural networks. [37], [39].
So far, a great number of research results concerning dis-
tributed state estimation have been available in the literature,
I. INTRODUCTION among which the most popular techniques are Kalman filter-
ing [14], [18], [20], [23], [30] and H∞ approach [7], [33], [37].
HE past decades have witnessed an exponential growth
T of the deployments of sensor networks in various fields
Note that the performance of Kalman filtering largely depends
on whether the noises meet the Gaussian assumption [28], [35],
while the H∞ approach is only applicable for handling the
Manuscript received April 13, 2021; revised June 16, 2021; accepted July so-called energy-bounded disturbances [21]. In many practical
9, 2021. Date of publication July 14, 2021; date of current version August
5, 2021. The associate editor coordinating the review of this manuscript and scenarios, unfortunately, neither Gaussian nor energy-bounded
approving it for publication was Dr. Lifeng Lai. This work was supported in assumptions could adequately reflect the noise properties. For in-
part by the Natural Science Foundation of Jiangsu Province of China under stance, in electronics and electrical engineering, the disturbance
Grant BK20190021, in part by the National Natural Science Foundation of
China under Grants 61773209, 61973102, 61873148, 61933007, and 71801196, caused by the man-made electromagnetic interference is more
in part by the Six Talent Peaks Project in Jiangsu Province of China under suitable to be modeled by the so-called unknown-but-bound
Grant XYDXX-033, and in part by the Alexander von Humboldt Foundation of (UBB) noise [1], [8]. By now, several approaches have been
Germany. (Corresponding author: Xiaojian Yi.)
Lifeng Ma is with the School of Automation, Nanjing University of Science exploited to study the state estimation issue against UBB noises
and Technology, Nanjing 210094, China (e-mail: [email protected]). (see e.g. [1], [3], [8], [16], [25], [47]), most of which are based
Zidong Wang is with the Department of Computer Science, Brunel on the linear matrix inequality approach and its variants in a
University London, Uxbridge, Middlesex UB8 3PH, U.K. (e-mail:
[email protected]). set-membership framework.
Yun Chen is with the School of Automation, Hangzhou Dianzi University, Aside from the exogenous noises, another factor that comes
Hangzhou 310018, China (e-mail: [email protected]). with the use of open yet unprotected communication networks
Xiaojian Yi is with the School of Mechatronical Engineering, Beijing Institute
of Technology, Beijing 100081, China (e-mail: [email protected]). is the malicious threats from adversaries. It has now been widely
Digital Object Identifier 10.1109/TSIPN.2021.3097217 recognized that cyber-attacks against the networks are prevalent
2373-776X © 2021 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: PAKISTAN INST OF ENGINEERING AND APPLIED SCIENCES. Downloaded on October 08,2022 at 17:10:56 UTC from IEEE Xplore. Restrictions apply.
 466 IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS, VOL. 7, 2021

that behave as a major source of the performance deterioration achieve desired performances with a satisfactory chance; and
or even the system crash [9], [43], [44]. In the context of iv) a joint estimation algorithm is provided to simultaneously
sensor networks, it is worth noting that a large part of malicious estimate the system state and falsified innovations.
attacks against the target systems are performed by injecting Notation Rn denotes the n-dimensional Euclidean space and
specific signals into the intercepted data during the information 1n denotes an n-dimensional column vector with all ones. In
transmission. By utilizing the falsified information, the adver- denotes the identity matrix of n dimensions. The notation A ≥ B
saries are able to deceive the receivers, thereby deteriorating or (respectively A > B), where A and B are symmetric matrices,
manipulating system performances [17]. As the reliability and means that A − B is positive semi-definite (respectively positive
safety are among the fundamental requirements of communi- definite). For matrices X and Y , their Kronecker product is
cation networks, the security-relevant issues have provoked an denoted as X ⊗ Y . For a vector x, x = xT x. For  a matrix
ever-growing attention, see, e.g. [17], [22], [42] for some recent Z, tr[Z] means the trace of matrix Z, and ZF = tr[Z T Z]
publications. Most existing results, however, have been based on is the Frobenius norm of Z. diag{X1 , X2 , . . . , Xn } denotes
the assumption that the falsified signals are known and linear, a block diagonal matrix whose diagonal blocks are given by
and such an assumption is often unrealistic in practical scenarios. X1 , X2 , . . . , Xn . The notation diagn {Xi } represents the block
As such, in this paper, we make dedicated efforts to handle diagonal matrix diag{X1 , X2 , . . . , Xn } and coln {xi } denotes
unknown deception attacks on the innovations by proposing a the column vector [xT T T T
1 x2 . . . xn ] . P {X} means the occur-
novel algorithm for the distributed state estimation problem with rence probability of the event ‘X’.
the aid of artificial neural networks (NN).
In practical engineering, it is often unnecessary (and unattain- II. PROBLEM FORMULATION
able) to design controllers/estimators which persistently achieve
The sensor network under consideration consists of N sensors
the desired performance indices with a 100% confidence level.
with the communication topology described by a directed graph
Instead, one would prefer to follow the so-called probability-
G = (V , E , L ), where V = {1, 2, ..., N } represents the set of
guaranteed design principle which aims at reaching the desired
sensing nodes, E ⊆ V × V represents the set of edges, and
indices with a satisfactory chance (less than 1) [36], [40]. Such
L = [θij ]N ×N represents the nonnegative adjacency matrix.
a concept could find wide applications in various engineering
Specifically, θij > 0 means that sensor j can send information
branches, for example, the target tracking issue and weapon
to sensor i (in such a case, j is called a neighbour of i), while
shooting tests [27]. The advantage of such a concept lies in its
θij = 0 indicates there are no data transmitted from node j to
capability of ensuring satisfactory performance while i) avoiding
node i. Denote by Ni  {j ∈ V |(i, j) ∈ E } the set of neigh-
unnecessarily stringent requirements and ii) leaving additional
bours of node i. Moreover, we assume that θii = 0 for all i ∈ V .
design freedom for other system specifications.
Consider the following nonlinear system defined on the hori-
In response to the above discussions, this paper aims to
zon [0, T ]:
develop a probability-guaranteed distributed state estimation ap-

proach over sensor networks where the communications among xk+1 = f (xk ) + Bk νk
nodes are under the threat of malicious manipulations. Artificial (1)
yi,k = gi (xk ) + Ei,k μk
neural networks are utilized to approximate the unknown fal-
sified innovations during data propagation among nodes. This where xk ∈ Rnx is the system state; yi,k ∈ Rny is the measure-
would be a very challenging problem due to some essential ment output of sensor node i; νk ∈ Rnν and μk ∈ Rnμ are the
difficulties summarized as follows. 1) For the unknown fal- process and measurement disturbances; Bk and Ei,k are known
sified innovations, it is invariably arduous to model the dy- real-valued matrices of compatible dimensions; f (xk ) : Rnx →
namical characteristics via a neural network and yet facilitate Rnx and gi (xk ) : Rnx → Rny are smooth nonlinear functions.
the subsequent development. 2) The cross coupling between Assumption 1: The noises νk and μk satisfy the following
the desired estimator parameters and the NN weight matrices conditions:
brings considerable difficulties in analysis and design. 3) The 
νk ∈ Vk  {νk : νkT Vk−1 νk ≤ 1}
joint estimation of the system state and falsified innovations (2)
−1
brings extra challenges when determining the estimator gains μk ∈ Uk  {μk : μT k Uk μk ≤ 1}
and NN weight matrices. Consequently, it is our main pur-
where Vk > 0 and Uk > 0 are known matrices with suitable
pose to tackle these challenges by investigating the addressed
dimensions.
probability-guaranteed secure distributed state estimation issue.
Before giving the structure of distributed NN-based state es-
The contributions of this paper can be identified as follows:i)
timator to be designed, we first recall the form of the traditional
the system under consideration is modeled by a nonlinear differ-
distributed estimators. For each node i (i = 1, 2, . . . , N ), the
ence equation, which is much more general than those studied in
local state estimator has the following form:
literature and could reflect engineering practice in a more com- 
prehensive way; ii) in order to depict appropriately dynamical x̂i,k+1 = Fi,k x̂i,k + θij Hij,k (yj,k − gj (x̂j,k ))
behaviors over a finite time interval of interest, a new transient j∈Ni
performance index is defined for system at each time step; iii) 
with the purpose of avoiding unnecessarily stringent require-  Fi,k x̂i,k + θij Hij,k zj,k (3)
ments, a probabilistic design principle is put forward aiming to j∈Ni

Authorized licensed use limited to: PAKISTAN INST OF ENGINEERING AND APPLIED SCIENCES. Downloaded on October 08,2022 at 17:10:56 UTC from IEEE Xplore. Restrictions apply.
 MA et al.: PROBABILITY-GUARANTEED DISTRIBUTED SECURE ESTIMATION FOR NONLINEAR SYSTEMS OVER SENSOR NETWORKS 467

where x̂i,k represents the state estimate of node i; zi,k  yi,k − paper, the local estimator on node i is reconstructed as follows:
ŷi,k is the innovation information with ŷi,k  gi (x̂i,k ) being the x̂i,k+1 = Fi,k x̂i,k + Gi,k (yi,k − ŷi,k )
estimated output; and the matrices Fi,k and Hij,k are estimator  

gains. + θij Hij,k z̃j,k − ᾱŴij,k φ(zi,k ) (8)
The most essential feature of a distributed estimation algo- j∈Ni
rithm is that each local estimator generates state estimates by
utilizing its own measurements in combination with neighbors’ where Fi,k , Gi,k and Hij,k are estimator parameters to be
information shared via communication networks according to designed; Ŵij,k is the estimate of the weight matrix Wij which
the interaction topology. Such an appealing feature would be is defined as the ideal weight matrix for Wj at sensing node i.
jeopardized with deteriorated estimation performance in the case Remark 2: It is worth noting that the innovations received
that the networks are under the threat of malicious manipula- by node i are not the original ones but containing the injected
tions, and this gives rise to the motivation for us to consider the signals, which implies that the true value of zj,k are not available
secure estimation issue in this paper. on the side of node i. Therefore, in (8), instead of zj,k , we
Consider the information propagation process from node j to shall utilize the local innovation zi,k as the input of the neural
node i. The adversary first intercepts the innovation information networks to approximate the injected signal χ(zj,k ). On the
sent by node j (i.e., zj,k ). Then, by using the intercepted zj,k , the other hand, it is obvious that, on the side of node i, we cannot
(a)
adversary generates a signal zj,k which will be utilized to inject obtain the exact value of the ideal weight matrix Wij due to the
(a) lack of original innovations sent from j. Therefore, in (8), we
into the original innovation. The injected signal zj,k is assumed use the estimated Ŵij,k to replace the ideal value of Wij , and
to have the following form: an associated updating law for Ŵij,k will be designed later in
(a)
zj,k = χ(zj,k ) (4) Subsection III-A for tuning Ŵij,k at each time step according to
the state estimation performance adaptively.
where χ(·) : Rny → Rny is an unknown nonlinear function Definition 1: An ellipsoid X(a, Y ) in Rn is defined by
defined on a compact set.
Taking (4) into account, the falsified innovation sent from X(a, Y )  {x ∈ Rn : (x − a)T Y −1 (x − a) ≤ 1} (9)
node j (after being manipulated) is described as follows: where a ∈ Rn represents the center and Y > 0 is a positive
(a) definite matrix with appropriate dimension.
z̃j,k = zj,k + αk zj,k = zj,k + αk χ(zj,k ) (5)
To give the design objective, we define the system state
where αk is a Bernoulli distributed random sequence satisfying estimation error by x̃i,k  xk − x̂i,k and the NN weights tuning
Prob{αk = 1} = ᾱ. Subsequently, the adversary will continue error by W̃ij,k  Wij − Ŵij,k .
to send such a manipulated message (5) to node i with the aim The following assumption concerning the initial conditions on
to deteriorate the system performance. the estimation errors will be needed for our further development.
Remark 1: It should be mentioned that, due to the complexity Assumption 3: The initial conditions of W̃ij,0 and x̃i,0 satisfy
of sensor networks, the adversary cannot perform the attacks  
T
successfully all the time but has certain limitations on ability. tr W̃ij,0 Q−1
ij,0 W̃ ij,0 ≤1
The Bernoulli distributed random variable αk is introduced to
describe the constraints or limitations imposed on the ability of (x0 − x̂i,0 )T P0−1 (x0 − x̂i,0 ) ≤ 1
the adversary. In other words, the innovation will be falsified
where Qij,0 and P0 are known positive definite real-valued
with a certain success rate.
matrices.
On the other hand, at node i, the local estimator tries to provide
It is the aim of this paper to design the estimator gains Fi,k ,
the state estimate by using local measurements as well as the
Gi,k and Hij,k in (8) such that the state estimation errors at
innovation information sent by its neighbours j ∈ Ni . Taking (5)
each sensing node satisfy the following probabilistic ellipsoidal
into account, we shall use the neural networks to approximate the
constraint:
unknown function χ(zj,k ) with the hope to mitigate the effects
from the falsified innovations as follows: P {xk ∈ X(x̂i,k , Pk )} ≥ p (10)
χ(zj,k ) = Wj φ(zj,k ) + δj,k (6) or, equivalently,
−1
where Wj is the ideal weight matrix of the neural network, φ(·) P {x̃T
i,k Pk x̃i,k ≤ 1} ≥ p (11)
is the activation function and δj,k is the approximation error.
Assumption 2: [38] The ideal weight matrix Wj , the activa- where Pk > 0 is a pre-specified matrix and p is a pre-specified
tion function φ(·) and the approximation error δj,k satisfy: scalar satisfying 0 < p < 1.

Wj F ≤ 1j , φ(·) ≤ 2 , δj,k  ≤ 3j (7) III. MAIN RESULTS

where 1j , 2 and 3j are known positive scalars. In this section, we will discuss the distributed state estimation
We are now in the position to present the structure of the local issue subject to falsified innovations. First, two useful lemmas
estimator by taking into account the falsified innovations. In this are introduced.

Authorized licensed use limited to: PAKISTAN INST OF ENGINEERING AND APPLIED SCIENCES. Downloaded on October 08,2022 at 17:10:56 UTC from IEEE Xplore. Restrictions apply.
 468 IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS, VOL. 7, 2021

Lemma 1: (S-procedure [2]) Define a series of quadratic func- fˆk  colN {f (x̂i,k )}, ĝk  colN {g(x̂i,k )}, δk  colN {δi,k },
tions κ0 (·), κ1 (·), . . . , κι (·) with respect to the variable a ∈ Rn
as κj (a)  aT Yj a where YjT = Yj (j = 0, . . . , ι). If there exist φk  colN {φ(zi,k )}, ηk  colN {yi,k }, Bk  1N ⊗ Bk ,
a sequence of non-negative scalars {1 , 2 , . . . , ι } satisfying
Gk  diagN {Gi,k }, Fk  diagN {Fi,k }, Ek  diagN {Ei,k },
Y0 − ιj=1 j Yj ≤ 0, then we have the following derivation:
W  [Wij ]N ×N , Ŵk  [Ŵij,k ]N ×N ,
κ1 (a) ≤ 0, . . . , κι (a) ≤ 0 → κ0 (a) ≤ 0.
L  diagN {Li }, Φk  diagN {Φi,k },
Lemma 2: (Schur Complement Equivalence) For the matri-
ces Y1 , Y2 , Y3 where Y1 = Y1T and 0 < Y2 = Y2T , the matrix Ψk  diagN {Ψi,k }, Σ  diagN {Σi },
inequality Y1 + Y3T Y2−1 Y3 < 0 if and only if
Δ1  diagN {Δ1i }, Δ2  diagN {Δ2i },
Y1 Y3T −Y2 Y3
< 0 or < 0. Θι,i  diag{0, . . . , 0, Iι , 0, . . . , 0},
Y3 −Y2 Y3T Y1
     
i−1 N −i
By using Taylor expansion technique, we describe f (xk ) and
gi (xk ) as follows: Rι,i  (1T
N ⊗ Iι )Θι,i , ι = {nx , ns , ny , nΣ }.
f (xk ) = f (x̂i,k ) + Φi,k x̃i,k + Li Δ1i x̃i,k (12) Then, we rewrite the estimation error dynamics in the following
form:
gi (xk ) = gi (x̂i,k ) + Ψi,k x̃i,k + Σi Δ2i x̃i,k (13)
x̃k+1 = fˆk + Φk x̃k + LΔ1 x̃k + Bk νk
where Li ∈ Rnl and Σi ∈ RnΣ are known matrices; Δ1i ∈
Rnl ×nx and Δ2i ∈ RnΣ ×nx are unknown matrices satisfying − Fk x̂k − Gk Ψk x̃k − Gk ΣΔ2 x̃k
Δ1i  ≤ 1 and Δ2i  ≤ 1; Φi,k and Ψi,k are obtained as fol-
− Gk Ek (1N ⊗ Inμ )μk − Hk Ψk x̃k − Hk ΣΔ2 x̃k
lows:
∂f (x) ∂gi (x) − Hk Ek (1N ⊗ Inμ )μk − αk (Hk ◦ W)φk
Φi,k  , Ψi,k  . (14)
∂x x=x̂i,k ∂x x=x̂i,k N

Remark 3: In (12) and (13), the terms Li Δ1i x̃i,k and + ᾱ Θnx ,i (Hk ◦ Ŵk )(1N ⊗ Rny ,i )φk
i=1
Σi Δ2i x̃i,k account for the truncation errors due to the ne-
glected higher order terms in the Taylor series expansion of = fˆk − Fk x̂k + (Φk + LΔ1 − Gk Ψk
the corresponding nonlinear functions. Alternatively, Δ1i and
−Gk ΣΔ2 − Hk Ψk − Hk ΣΔ2 ) x̃k
Δ2i can also be illustrated as uncertain terms that are employed
to describe the modeling errors in the matrices Φi,k and Ψi,k , + Bk νk − (Gk + Hk )Ek (1N ⊗ Inμ )μk
respectively. The readers are referred to [3] for a rigorous justifi-
− αk Hk δk − αk (Hk ◦ W)φk
cation of the validity of such a representation for the linearization
error. N

Then, the one-step estimation error dynamics are acquired as + ᾱ Θnx ,i (Hk ◦ Ŵk )(1N ⊗ Rny ,i )φk (16)
follows: i=1

x̃i,k+1 = xk+1 − x̂i,k+1 where Hk  [θij Hij,k ]N ×N . Note that since θij = 0 when j ∈
/
Ni , Hk is a sparse matrix described by
= f (xk ) + Bk νk
Hk ∈ Tnx ×ny
− (Fi,k x̂i,k + Gi,k (yi,k − ŷi,k )
⎞ where Tnx ×ny  {T = [Tij ] ∈ RN nx ×N ny Tij ∈ Rnx ×ny ,
   Tij = 0 if j ∈
/ Ni }. Moreover, in equation (16), for two block
+ θij Hij,k z̃j,k − ᾱŴij,k φ(zi,k ) ⎠ matrices A = [Aij ]N ×N and B = [Bij ]N ×N where Aij and
j∈Ni
Bij are matrices with compatible dimensions, the product of
= f (x̂i,k ) + Φi,k x̃i,k + Li Δ1i x̃i,k + Bk νk A ◦ B is defined by A ◦ B  [Aij Bij ]N ×N .
− (Fi,k x̂i,k + Gi,k (Ψi,k x̃i,k + Σi Δ2i x̃i,k + Ei,k μk )
A. Tuning of Ŵij,k

+ θij Hij,k (Ψj,k x̃j,k + Σj Δ2j x̃j,k + Ej,k μk The adaptive tuning law of the estimated weight matrix Ŵij,k
j∈Ni is designed as follows:

+αk Wij φ(zj,k ) + αk δj,k − ᾱŴij,k φ(zi,k ) . (1) (2)
Ŵij,k+1 = λij,k Ŵij,k + λij,k k φT (zi,k ) (17)
(15) where

For brevity of development, we define k  ᾱ T
θij Hij,k ΨTi,k+1 zi,k+1
ξk  colN {xk }, x̂k  colN {x̂i,k }, x̃k  colN {x̃i,k }, j∈Ni

Authorized licensed use limited to: PAKISTAN INST OF ENGINEERING AND APPLIED SCIENCES. Downloaded on October 08,2022 at 17:10:56 UTC from IEEE Xplore. Restrictions apply.
 MA et al.: PROBABILITY-GUARANTEED DISTRIBUTED SECURE ESTIMATION FOR NONLINEAR SYSTEMS OVER SENSOR NETWORKS 469

(1) (2) is true, then we have

and λij,k and λij,k are positive tuning scalars that will be
determined later. P {υ ∈ X(a, Y )} ≥ p.
Remark 4: The adaptive tuning law (17) is proposed ac-
cording to the gradient descent algorithm aiming to minimize Define a matrix Pk by
the cost function defined by J  zi,k+1  = yi,k+1 − ŷi,k+1 . 1
Pk  P ,
(1) (2)
The tuning scalars λij,k and λij,k will be co-designed with the 1−p k
parameters of state estimators subsequently. and then, on basis of Lemma 3, we immediately obtain the
The following theorem gives a sufficient condition under following lemma.
which the matrix W̃ij,k will be bounded in the sense of weighted Lemma 4: If E{(xk − x̂i,k )T Pk−1 (xk − x̂i,k )} ≤ 1, then the
Frobenius norm. following holds:
Theorem 1: For nonlinear system (1), the tuning law
of Ŵij,k is given by (17). Let the family of matrices P {xk ∈ X(x̂i,k , Pk )} ≥ p. (20)
{Qij,k }k∈[0,T ] be given. If there exist families of positive Proof: Lemma 4 is easily proved according to Lemma 3, and
(1) (2)
scalars {ij,k , ij,k }k∈[0,T −1] and families of tuning scalars the proof is therefore omitted here. 
(1) (2)
{λij,k , λij,k }k∈[0,T −1] satisfying: Defining α̃k  αk − ᾱ, we know that
E{α̃k } = 0, E{α̃k2 } = ᾱ(1 − ᾱ)  σ 2 . (21)
Ωij,k ∗
≤0 (18)
Υij,k −Qij,k+1 Theorem 2: Let the estimator gains Fi,k , Gi,k and Hij,k be
given. Under the condition of Theorem 1, for a pre-specified
where
family of positive definite matrices {Pk }k∈[0,T ] with a factor-
Qij,k  Qij,k ⊗ Iny , ization of Pk = Sk SkT , if there exist families of non-negative
(1) (2) (3) (4) (5) (6) (7)
(1) (1) (1) (2) scalars {βi,k , βk , βk , βk , βi,k , βi,k , βi,k }k∈[0,T −1] satisfy-
Ωij,k  diag{−1 + ij,k + 21j ij,k , −ij,k I, −ij,k I},
ing the following recursive matrix inequalities:
  ⎡ ⎤
Υij,k  −λ(2) 
ij,k k ⊗ φ(z i,k ) λ
(1)
M
ij,k ij,k (1 − λ
(1)
ij,k )I −Γk ∗ ∗
⎣ Rnx ,i Ξ̄k −Pk+1 ∗ ⎦≤0 (22)
with Mij,k being a factorization of the matrix Qij,k (i.e., Qij,k =
T σRnx ,i Ξ̃k 0 −Pk+1
Mij,k Mij,k ), then the following inequality holds:
where
T
tr[W̃ij,k+1 Q−1
ij,k+1 W̃ij,k+1 ] ≤ 1. (19)  N
 (1) (5) (2) (3) (4)
Proof: See Appendix A.  Γk  diag 1− (βi,k + βi,k 3i ) − βk − βk −N 2 βk ,
Remark 5: Theorem 1 gives a sufficient condition to con- i=1

strain the estimation errors of NN weight matrix (i.e., W̃ij,k ) N

 N

(1) (6) (7)
to be bounded by a pre-specified criterion (i.e.,Qij,k ) in the βi,k Θns ,i − (βi,k + βi,k )Θns ,i SkT Sk ,
sense of weighted Frobenius norm. From (17) we know that i=1 i=1
the matrix Ŵij,k+1 will be tuned adaptively by making the N
 N

(6) (7) (2) (3)
trade-off between the current step weight matrix (i.e., Ŵij,k ) βi,k Θnl ,i , βi,k ΘnΣ ,i , βk Vk−1 , βk Uk−1 ,
and the innovation information (i.e., zi,k+1 ). This theorem will i=1 i=1
play a paramount role in the following analysis and design of N

 (5) (4)
the desired state estimator gains. The connection between the βi,k Θnδ ,i , βk IN 2 , (23)
tuning errors of NN weight and the estimation errors of system i=1
state will be discussed in detail in Remark 6 of Subsection III-C. 
(11) (12)
Ξ̄k  Ξ̄k Ξk L −(Gk + Hk )Σ
B. Ellipsoidal Constraint in Probability p 
(16)
Bk Ξk −ᾱHk −ᾱHk , (24)
In this subsection, we shall establish the sufficient condition
under which the state estimation error will be driven to reside (11)
Ξ̄k  fˆk − Fk x̂k − ᾱ(Hk ◦ Ŵk )φk
within an allowable ellipsoidal area in a pre-specified probabil-
ity. First, we introduce the following lemmas that will be used N

in the derivation. + ᾱ Θnx ,i (Hk ◦ Ŵk )(1N ⊗ Rny ,i )φk , (25)
Lemma 3: [36] Given a random variable υ of suitable dimen- i=1
sion which belongs to an ellipsoid X(a, Y ) as follows: 
Ξ̃k  −(Hk ◦ Ŵk )φk 0 0 0
 
υ ∈ X(a, Y )  υ|(υ − a)T Y −1 (υ − a) ≤ 1 
0 0 −Hk −Hk , (26)
where a and Y are given in Definition 1. If, for any given 0 < (12)
p < 1, the following inequality Ξk  (Φk − Gk Ψk − Hk Ψk )Sk , (27)
  (16)
E (υ − a)T Y −1 (υ − a) ≤ 1 − p Ξk  − (Gk + Hk )Ek (1N ⊗ Inμ ), (28)

Authorized licensed use limited to: PAKISTAN INST OF ENGINEERING AND APPLIED SCIENCES. Downloaded on October 08,2022 at 17:10:56 UTC from IEEE Xplore. Restrictions apply.
 470 IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS, VOL. 7, 2021

then the following inequality holds: the computational complexity of our proposed RLMI algorithm
depends linearly on the length of time interval T + 1. It is worth
E{(xk+1 − x̂i,k+1 )T Pk+1
−1
(xk+1 − x̂i,k+1 )} ≤ 1. (29) mentioning that the study on LMI optimization is very active
Proof: See Appendix B.  in recent years within the communities of applied mathematics,
control science and signal processing. We can expect substantial
speedups in the near future.
C. Distributed Estimator Design
The following algorithm provides a computing method to it-
In this subsection, a sufficient condition is provided for the eratively obtain the required estimator gains {Fi,k , Gi,k , Hij,k }.
solvability of the addressed probability-guaranteed distributed
state estimator design problem. The desired estimator gains at Algorithm 1: Distributed Estimator Design Algorithm.
each time step can be acquired by solving the corresponding
matrix inequalities. 1: Input: Topology information G = (V , E , L ); Initial
Theorem 3: Given a pre-specified 0 < p < 1 and a fam- values: xi,0 , x̂i,0 , Wij,0 , Ŵij,0 , Qij,0 , and P0 ;
ily of positive definite matrices {Pk }k∈[0,T ] . Under the Performance constraints {Qij,k }k>0 , {Pk }k>0 and p;
condition given in Theorem 1, the design objective (11) Activation function φ(·); Other coefficients:
is satisfied if there exist families of non-negative scalars ᾱ, 1i , 2 , 3i , Uk and Vk ; Maximum time step T .
(1) (2) (3) (4) (5) (6) (7) 2: Set k = 0
{βi,k , βk , βk , βi,k , βi,k , βi,k , βi,k }k∈[0,T −1] such that
3: if k < T then
⎡ ⎤ 4: Solve (30) for estimator gains {Fi,k , Gi,k , Hij,k }.
−Γk ∗ ∗
1
⎣ Rnx ,i Ξ̄k − 1−p Pk+1 ∗ ⎦ ≤ 0. 5: Obtain one-step estimate x̂i,k+1 according to (8).
(30)
1 6: Obtain innovation zi,k+1 .
σRnx ,i Ξ̃k 0 − 1−p Pk+1 (1) (2)
7: Solve (18) for tuning coefficients λij,k and λij,k .
Moreover, the estimator gains Fi,k , Gi,k and Hij,k can be 8: Obtain Ŵij,k+1 according to (17).
calculated via solving the corresponding matrix inequalities. 9: end if
Proof: Theorem 3 can be proved easily from Lemma 4 and
1
Theorem 2 by taking Pk+1 = 1−p Pk+1 into account, and thus
is omitted here. 
D. Optimization Problems
Remark 6: Theorem 3 provides a scheme to obtain the desired
estimator guaranteeing the state estimation errors are confined It should be noticed that within the design framework we
within an allowable ellipsoidal area in probability p. Note that proposed in Subsection III-C, the acquired estimator parameters
Theorem 3 has an inherent connection with Theorem 1. On one could be a set if existing. Therefore, in this subsection, we aim
hand, the boundedness of W̃ij,k , ensured by Theorem 1, is of help to present the optimization problems to seek the locally opti-
to restrict the state estimation error x̃i,k ; on the other hand, the mal estimator parameters by making use of trade-off between
satisfaction of the requirement imposed on the state estimation performance specifications.
errors, guaranteed by Theorem 3, contributes to constrain W̃ij,k . OP1: Minimization of Qij,k (in the sense of matrix trace) to
In specific, Theorem 1 influences the state estimation via (37) ensure the minimal estimation errors of the NN weight matrices
and (53), while Theorem 3 helps to regulate the tuning of Ŵij,k (in the sense of weighted Frobenius norm)
via state estimate x̂i,k and the resulting innovation characterized Corollary 1: Under the conditions given in Theorem 1, a
by formula (17). In such a coupling yet collaborative way, the family of minimized matrices {Qij,k }k∈[0,T ] can be obtained in
desired state estimation performance can be reached. the sense of matrix trace if the following minimization problem
Remark 7: Notice that the RLMI algorithm proposed in is feasible:
this paper is based on LMI approach. As discussed in [2], min tr[Qij,k+1 ] (31)
the computational complexity of an LMI system is bounded (1) (2) (1) (2)
{ij,k ,ij,k ,λij,k ,λij,k ,Qij,k+1 }
by O(PQ3 log(U /ε)) where P represents the row size, Q
stands for the number of scalar decision variables, U is a
data-dependent scaling factor and ε is relative accuracy set subject to (18)
for algorithm. For instance, let us now look at the condition In the following, on basis of Corollary 1, we shall further
proposed in Theorem 1, where the number of sensing nodes is present two optimization problems to guarantee locally optimal
N , the iteration time is T + 1 (since the time interval is [0, T ]) state estimation performance from different perspectives. To this
and the dimensions of variables are known from xk ∈ Rnx , end, for brevity of development, denote a set of variables by
yi,k ∈ Rny , νk ∈ Rnν and μk ∈ Rnμ . Moreover, we assume that (1) (2) (3) (4) (5) (6) (7)
ri,k ∈ Rnr and χ(zj,k ) ∈ Rnχ . The RLMI-based algorithm is Sk  {Fk , Gk , Hk , βi,k , βk , βk , βk , βi,k , βi,k , βi,k }.
implemented recursively for T + 1 steps and, at each step, we OP2: Minimization of Pk (in the sense of matrix trace) to
need to solve the LMI (18) with P = N 2 (1 + nr + nχ ṅy ) rows guarantee the locally optimal estimation subject to the fixed pre-
n2 +n
and Q = N 2 ( χ 2 χ + 4) scalar variables. Accordingly, the specified probability specification.
computational complexity of the proposed RLMI algorithm can Corollary 2: Let p be given. Under the conditions given in
be represented by O((T + 1)PQ). We can now conclude that Theorem 2 and Corollary 1, a family of minimized matrices

Authorized licensed use limited to: PAKISTAN INST OF ENGINEERING AND APPLIED SCIENCES. Downloaded on October 08,2022 at 17:10:56 UTC from IEEE Xplore. Restrictions apply.
 MA et al.: PROBABILITY-GUARANTEED DISTRIBUTED SECURE ESTIMATION FOR NONLINEAR SYSTEMS OVER SENSOR NETWORKS 471

{Pk }k∈[0,T ] can be ensured in the sense of matrix trace if the IV. A NUMERICAL EXAMPLE
following minimization problem is feasible: In this section, an illustrative example is presented to demon-
min tr[Pk+1 ] (32) strate the usefulness of the proposed algorithm in this paper.
{Sk ,Pk+1 }
First, for a vector c ∈ R2 , denote by c(1) and c(2) the first and
second entries of c, respectively.
subject to (22) Consider a nonlinear system whose parameters are given as
follows:
Next, suppose p is time-varying and denote by pk the prob-
ability criterion at instant k. Define 0.1 + 0.01 cos(k + 1)
Bk = ,
0.2 + 0.01 sin(k)
1
qk  .
1 − pk E1,k =
0.25
,
0.2 + 0.1 cos(k + 1)
OP3: Minimization of qk subject to fixed Pk to determine
the lower bound on the probability criterion at each time step. 0.2
Corollary 3: Let {Pk }k∈[0,T ] be given. Under the conditions E2,k = ,
0.15 + 0.1 sin(2k)
in Theorem 3 and Corollary 1, the lower bound on pk is guar-
0.25 + 0.05 sin(k)
anteed if the following minimization problem is feasible: E3,k = ,
0.15
min qk (33) !
{Sk ,qk } (1) (1) (2)
0.5xk + 0.45 sin(xk ) + 0.25xk
⎧ f (xk ) = (2) (2) (1) ,
⎪ 1 < qk < +∞ 0.4 cos(xk ) + 0.3xk + 0.6xk
⎪
⎪
⎪
⎪ ⎡ ⎤ !
⎪
⎨ −Γk ∗ ∗ (1) (1) (2)
0.15xk + 0.35 sin(2xk ) + 0.2xk
subject to ⎣ Rnx ,i Ξ̄k −qk Pk+1 ∗ ⎦≤0 g1 (xk ) = (2) (2) (1) ,
⎪
⎪ 0.5 sin(xk ) + 0.1xk + 0.3xk
⎪
⎪ σRnx ,i Ξ̃k 0 −qk Pk+1
⎪
⎪ !
⎩ (1) (1) (2)
0.3xk + 0.3 sin(xk + 1) + 0.4xk
g2 (xk ) = (2) (2) (1) ,
The proofs of Corollaries 1–3 are quite simple on basis of the 0.45 cos(xk ) + 0.25xk + 0.5xk
obtained theorems, and are thus omitted here. !
(1) (1) (2)
0.5 sin(xk ) + 0.4xk + 0.2xk
Remark 8: So far, we have discussed the probability- g3 (xk ) = (2) (2) (1) .
guaranteed distributed state estimation problem for nonlinear 0.4 sin(xk ) + 0.1xk + 0.3xk
system over sensor networks under random falsified innova-
We select νk = 0.3 cos(2k) and μ(k) = 0.4 sin(k). Set V =
tions. The solvability of the addressed problem is cast into the
1 and U = 1. It can be easily verified that Assumption 1 is
feasibility of a set of recursive linear matrix inequalities. A
satisfied.
methodology has been proposed which utilizes neural networks
Set 1i = 2 = 3i = 3. Choose the activation function of the
to approximate the unknown falsified innovations. Within the
neural networks as follows:
established framework, the tuning parameters of NN weight !
(1)
matrices and the probabilistic ellipsoidal constraint can be in- tanh(zi,k )
vestigated systematically. The desired estimator gains ensuring φ(zi,k ) = (2) .
tanh(zi,k )
required performance criterion can be obtained by solving the
recursive linear matrix inequalities at each time step. Three Assume that the number of sensor nodes is three, and the
optimization problems are proposed to ensure the sub-optimal adjacency matrix L is set as follows:
estimation performances from different perspectives. It should ⎡ ⎤
010
be emphasized that, it is not difficult to extend our obtained re- L = ⎣1 0 1⎦.
sults to the case where more performance indices are considered 010
simultaneously within the proposed theoretical framework.
Remark 9: Our proposed algorithm possesses the following Set p = 0.8. The initial values in this simulation are selected
merits: i) the recursive nature of the algorithm is a true reflection as follows:
of the time-varying feature of the addressed estimation error 2.0 1.5
dynamics; ii) the idea of using NN to approximate the unknown x̂1,0 = , x̂2,0 = ,
1.4 1.8
falsified innovations is of more practical significance in com-
parison to those existing methods where the attack functions are 1.7 1.8
x̂3,0 = , x0 = ,
usually assumed to be known to the defenders; and iii) within 1.9 1.5
the established generic framework, it is not difficult to apply
0.6887 −0.0588
our developed methodology to deal with systems with more P0 = ,
−0.0588 0.7390
complex dynamics such as quantization effects and Markovian
jump parameters, or handle more performance indices such as 10 20 0
Ŵij,0 = , Qij,0 = .
H∞ specification and robustness. 01 0 20

Authorized licensed use limited to: PAKISTAN INST OF ENGINEERING AND APPLIED SCIENCES. Downloaded on October 08,2022 at 17:10:56 UTC from IEEE Xplore. Restrictions apply.
 472 IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS, VOL. 7, 2021

(1)
Fig. 1.
(1) (1)
Trajectories of xk and x̂i,k (by estimator with NN). Fig. 3. x̃i,k (by estimator with NN).

(2)
(2) (2)
Fig. 4. x̃i,k (by estimator with NN).
Fig. 2. Trajectories of xk and x̂i,k (by estimator with NN).

The random falsified innovations are governed by the follow-

ing parameters:
!
(1)
0.15 sin(zi,k )
ᾱ = 0.6, χ(zi,k ) = (2) .
0.30 cos(zi,k )
By solving the optimization problems in Corollaries 1–3,
the simulation results are obtained in Figs. 1– 4. Specifically,
(1)
Figs. 1–2 depict the trajectories of the system state xk (i.e., xk
(2)
and xk ) and the corresponding estimates. Figs. 3–4 present the
(1) (2)
trajectories of the estimation errors x̃i,k and x̃i,k , respectively. It
can be seen from Fig. 1 and Fig. 2 that the developed technique
can estimate the entries of state (i.e., xk and xk ) with a satisfac- Fig. 5.
(1) (1)
Trajectories of xk and x̂i,k (by conventional estimator).
tory precision, which is also demonstrated via Fig. 3 and Fig. 4
where both the estimation errors are confined within the required
range. From the simulation figures we can clearly observe that A comparative simulation is carried out where we apply the con-
the proposed algorithm can effectively provide state estimates ventional algorithm (3) to generate local state estimates under the
on each sensor node. identical manipulations of innovations. The simulation results
Next, we proceed to show the superiority of the proposed by using the conventional algorithm are recorded in Figs. 5– 8.
algorithm on mitigating the influence from falsified innovations In specific, Figs. 5–6 plot the trajectories of the system state xk
(1) (2)
over the conventional distributed estimator characterized by (3). (i.e., xk and xk ) and their estimates, respectively. Figs. 7–8

Authorized licensed use limited to: PAKISTAN INST OF ENGINEERING AND APPLIED SCIENCES. Downloaded on October 08,2022 at 17:10:56 UTC from IEEE Xplore. Restrictions apply.
 MA et al.: PROBABILITY-GUARANTEED DISTRIBUTED SECURE ESTIMATION FOR NONLINEAR SYSTEMS OVER SENSOR NETWORKS 473

approach, the conventional algorithm shows worse estimation

performance in the existence of falsified innovations.

V. CONCLUSION
In this paper, we have investigated the distributed state esti-
mation problem for a class of general nonlinear systems subject
to falsified innovations. A neural-network-based mechanism has
been proposed to approximate the unknown manipulated data,
thereby mitigating the effects on the state estimation. With the
help of certain convex optimization methods, sufficient condi-
tions have been established for the solvability of the addressed
estimator design issue, ensuring that the state estimation errors
at each sensing node reside within required ellipsoidal regions in
a prescribed probability. The desired estimator gains have been
(2) (2) given in terms of the solution to a set of recursive matrix in-
Fig. 6. Trajectories of xk and x̂i,k (by conventional estimator).
equalities. On basis of the established framework, optimization
problems have been considered with the aim to ensure locally
optimal estimation performance. Finally, an illustrative numeri-
cal example has been presented to demonstrate the effectiveness
of the obtained theoretical results.

APPENDIX
A. Proof of Theorem 1
Proof: The proof is carried out by mathematical induction.
First, according to Assumption 3, at time step k = 0,
T
tr[W̃ij,0 Q−1
ij,0 W̃ij,0 ] ≤ 1 holds.
Second, suppose at time step k, tr[W̃ij,kT
Q−1
ij,k W̃ij,k ] ≤ 1 is
true. Then, we only need to verify that at time step k + 1, under
the condition given in this theorem, the following inequality also
holds:
T
Fig. 7.
(1)
x̃i,k (by conventional estimator). tr[W̃ij,k+1 Q−1
ij,k+1 W̃ij,k+1 ] ≤ 1. (34)

To this end, it is readily from the tuning law (17) that

W̃ij,k+1 = Wij − Ŵij,k+1

 
(1) (2)
= Wij − λij,k Ŵij,k + λij,k k φT (zi,k )
(1) (1) (2)
= (1 − λij,k )Wij + λij,k W̃ij,k − λij,k k φT (zi,k ).
(35)

Define a function ϕ(·) : Rm×n → Rmn as follows:

 T
ϕ(A)  A(1) A(2) · · · A(m)

where A(ι) stands for the ι-th row of matrix A. Then, by denoting
 T
(1) (2) (n )
ϕ(W̃ij,k )  W̃ij,k W̃ij,k · · · W̃ij,ky ,
(2)
Fig. 8. x̃i,k (by conventional estimator). we know that the following inequality
T
(1) (2)
tr[W̃ij,k Q−1
ij,k W̃ij,k ] ≤ 1
depict the evolutions of the estimation errors x̃i,k and x̃i,k ,
respectively. From the figures we can see that the conventional can be equivalently expressed by
algorithm is also capable of providing state estimation with
certain precision. However, in comparison to our developed ϕT (W̃ij,k )Q−1
ij,k ϕ(W̃ij,k ) ≤ 1. (36)

Authorized licensed use limited to: PAKISTAN INST OF ENGINEERING AND APPLIED SCIENCES. Downloaded on October 08,2022 at 17:10:56 UTC from IEEE Xplore. Restrictions apply.
 474 IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS, VOL. 7, 2021

Consequently, according to [3], taking into account Qij,k = Then, on basis of the principle of mathematical induction, the
T
Mij,k Mij,k , we have rest of the proof is to demonstrate that, under the given condition,
inequality (29) also holds at time step k + 1. In fact, since (49) is
ϕ(W̃ij,k ) = Mij,k rij,k (37) true, we can always find a vector si,k ∈ Rns (i = 1, 2, . . . , N )
where rij,k satisfies with E{sT i,k si,k } ≤ 1 such that
T
rij,k rij,k ≤ 1. (38) xk = x̂i,k + Sk si,k . (50)
By defining a vector πij,k as follows: By denoting sk  colN {si,k } and Sk  diagN {Sk }, (50) is
 T T described by
πij,k  1 rij,k ϕT (Wij ) , (39)
ξk = x̂k + Sk sk . (51)
we rewrite the error dynamics (35) by
Hence, (16) is rewritten by
ϕ(W̃ij,k+1 ) = Υij,k πij,k . (40)
x̃k+1 = fˆk − Fk x̂k + (Φk − Gk Ψk − Hk Ψk )Sk sk
T
Next, in terms of πij,k , the inequality rij,k rij,k ≤ 1 can be
+ Lρ1k − Gk Σρ2k − Hk Σρ2k + Bk νk
reformulated by
T
πij,k M̄ij,k πij,k ≤ 0 (41) − (Gk + Hk )Ek (1N ⊗ Inμ )μk − αk Hk δk

where M̄ij,k  diag{−1, I, 0}. − αk Hk ◦ W̃k φk − αk Hk ◦ Ŵk φk

Similarly, based on Assumption 2, one has N

+ ᾱ Θnx ,i Hk ◦ Ŵk (1N ⊗ Rny ,i )φk
tr[WijT Wij ] ≤ 21j , (42)
i=1
which, in terms of πij,k , is equivalently described by
= fˆk − Fk x̂k + (Φk − Gk Ψk − Hk Ψk )Sk sk
T
πij,k W̄ij,k πij,k ≤ 0 (43)
+ Lρ1k − Gk Σρ2k − Hk Σρ2k + Bk νk
where W̄ij,k  diag{−21j , 0, I}. − (Gk + Hk )Ek (1N ⊗ Inμ )μk − (ᾱ + α̃k )Hk δk
On the other hand, it is easily inferred that
T − (ᾱ + α̃k )(Hk ◦ W̃k )φk − (ᾱ + α̃k )(Hk ◦ Ŵk )φk
tr[W̃ij,k+1 Q−1
ij,k+1 W̃ij,k+1 ] ≤ 1 (44)
N

is equivalent to
+ ᾱ Θnx ,i (Hk ◦ Ŵk )(1N ⊗ Rny ,i )φk
ϕT (W̃ij,k+1 )Q−1
ij,k+1 ϕ(W̃ij,k+1 ) ≤ 1, (45) i=1

which can be further expressed as follows: = fˆk − Fk x̂k + (Φk − Gk Ψk − Hk Ψk )Sk sk

 
T
πij,k ΥT Q −1
Υ ij,k − diag{1, 0, 0} πij,k ≤ 0. (46) + Lρ1k − Gk Σρ2k − Hk Σρ2k + Bk νk
ij,k ij,k+1

Therefore, we know from Lemma 1 (S-procedure) that, if there − (Gk + Hk )Ek (1N ⊗ Inμ )μk − (ᾱ + α̃k )Hk δk
(1) (2)
exist positive scalars ij,k and ij,k such that the following N

inequality holds: − (ᾱ + α̃k ) Θnx ,i Hk W̃i,k φk
i=1
ΥT −1
ij,k Qij,k+1 Υi,k − diag{1, 0, 0}
− (ᾱ + α̃k )(Hk ◦ Ŵk )φk
(1) (2)
−ij,k M̄ij,k − ij,k W̄ij,k ≤ 0, (47) N

then inequality (45) also holds. + ᾱ Θnx ,i (Hk ◦ Ŵk )(1N ⊗ Rny ,i )φk (52)
Finally, according to Lemma 2 (Schur Complement Equiva- i=1
lence), we derive that (47) holds if and only if (18) holds. The where
proof is complete now. 
W̃i,k  diag{W̃i1,k , W̃i2,k , . . . , W̃iN,k },
B. Proof of Theorem 2 ρ1k  colN {ρ1i,k } = Δ1 Sk sk ,
Proof: The proof is performed by resorting to mathematical
ρ2k  colN {ρ2i,k } = Δ2 Sk sk .
induction approach. First, it is readily inferred from Assumption
3 that It is known that, under the condition of Theorem 1, we acquire
E{x̃T −1
i,0 P0 x̃i,0 } ≤ 1. (48) ϕ(W̃ij,k ) = Mij,k rij,k (53)
Second, suppose that at step k > 0, the following inequality T
with vector rij,k satisfying rij,k rij,k ≤ 1.
holds: By defining
E{x̃T −1
i,k Pk x̃i,k } ≤ 1. (49) φ̃k  Iny ⊗ φT (zi,k ),

Authorized licensed use limited to: PAKISTAN INST OF ENGINEERING AND APPLIED SCIENCES. Downloaded on October 08,2022 at 17:10:56 UTC from IEEE Xplore. Restrictions apply.
 MA et al.: PROBABILITY-GUARANTEED DISTRIBUTED SECURE ESTIMATION FOR NONLINEAR SYSTEMS OVER SENSOR NETWORKS 475

we have ζkT diag{−1, 0, 0, 0, Vk−1 , 0, 0, 0}ζk ≤ 0,

W̃ij,k φi = φ̃k ϕ(W̃ij,k ) = φ̃k Mij,k rij,k . ζkT diag{−1, 0, 0, 0, 0, Uk−1 , 0, 0}ζk ≤ 0,
Therefore, one obtains
ζkT diag{−N 2 , 0, 0, 0, 0, 0, 0, IN 2 }ζk ≤ 0,
W̃i,k φk = diagN {φ̃k } · diag{Mi1,k , Mi2,k , . . . , MiN,k } · ri,k
ζkT diag{−3i , 0, 0, 0, 0, 0, Θnδ ,i , 0}ζk ≤ 0.
 Mi,k ri,k
T T T
where ri,k  [ ri1,k ri2,k · · · riN,k ]T .
Moreover, from
By defining a new vector ζk as
 
T T T T T Δ1i  ≤ 1 and Δ2i  ≤ 1
ζk  1 s T T T
k ρ1k ρ2k νk μk δk rk
T T T we have
where rk  [ r1,k r2,k · · · rN,k ]T , the dynamics of state esti-
mation error are formulated as follows: ρT T T
1i,k ρ1i,k − si,k Sk Sk si,k ≤ 0,

x̃k+1 = Ξk ζk  (Ξ̄k + α̃k Ξ̃k )ζk (54) ρT T T

2i,k ρ2i,k − si,k Sk Sk si,k ≤ 0,
where which can be equivalently described by

Ξk  (11)
Ξk
(12)
Ξk L −(Gk + Hk )Σ Bk  
ζkT diag 0, −Θns ,i SkT Sk , Θnl ,i , 0, 0, 0, 0, 0 ζk ≤ 0,
  
Ξk
(16)
−(ᾱ + α̃k )Hk −(ᾱ + α̃k )Hk , ζkT diag 0, −Θns ,i SkT Sk , 0, ΘnΣ ,i , 0, 0, 0, 0 ζk ≤ 0.
(11)
Ξk  fˆk − Fk x̂k − (ᾱ + α̃k )(Hk ◦ Ŵk )φk
N Next, by virtue of Lemma 2 (Schur Complement Equiva-

+ ᾱ Θnx ,i (Hk ◦ Ŵk )(1N ⊗ Rny ,i )φk , lence), we know that inequalities (22) hold if and only if
i=1 Ξ̄T T −1
k Rnx ,i Pk+1 Rnx ,i Ξ̄k
(12)
Ξk  (Φk − Gk Ψk − Hk Ψk )Sk , +σ 2 Ξ̃T T −1
k Rnx ,i Pk+1 Rnx ,i Ξ̃k − Γk ≤ 0, (55)
(16)
Ξk  − (Gk + Hk )Ek (1N ⊗ Inμ ), which, by considering the statistical property of α̃k (i.e., formula
 (21)), implies that
Ξ̄k  Ξ̄(11)
k
(12)
Ξk L −(Gk + Hk )Σ Bk
 E{(Ξ̄k + α̃k Ξ̃k )T RT −1
nx ,i Pk+1 Rnx ,i (Ξ̄k + α̃k Ξ̃k )} − Γk ≤ 0.
(16)
Ξk −ᾱHk −ᾱHk , (56)
(11) By taking (23) and (54) into account, inequality (56) is equiv-
Ξ̄k  fˆk − Fk x̂k − ᾱ(Hk ◦ Ŵk )φk
alent to
N

+ ᾱ Θnx ,i (Hk ◦ Ŵk )(1N ⊗ Rny ,i )φk , E{x̃T −1 T
i,k+1 Pk+1 x̃i,k+1 } − ζk diag{1, 0, 0, 0, 0, 0, 0, 0}ζk
i=1 N
   (1)
Ξ̃k  − (Hk ◦ Ŵk )φk 0 0 0 0 0 Hk Hk , − βi,k ζkT diag{−1, Θns ,i , 0, 0, 0, 0, 0, 0}ζk
i=1
Hi,k  Θnx ,i Hk Mi,k , (2)
  −βk ζkT diag{−1, 0, 0, 0, Vk−1 , 0, 0, 0}ζk
Hk  H1,k H2,k · · · HN,k . (3)
−βk ζkT diag{−1, 0, 0, 0, 0, Uk−1 , 0, 0}ζk
It is already known that the following conditions hold:
(4)
−βk ζkT diag{−N 2 , 0, 0, 0, 0, 0, 0, IN 2 }ζk
E{sT
i,k si,k } ≤ 1,
N

νkT Vk−1 νk ≤ 1, −
(5)
βi,k ζkT diag{−3i , 0, 0, 0, 0, 0, Θnδ ,i , 0}ζk
i=1
μT −1
k Uk μk ≤ 1,
N

T
rij,k rij,k ≤ 1, −
(6)
βi,k ζkT diag{0, −Θns ,i SkT Sk , Θnl ,i , 0, 0, 0, 0, 0}ζk
i=1
δi  ≤ 3i ,
N
 (7)
− βi,k ζkT diag{0, −Θns ,i SkT Sk , 0, ΘnΣ ,i , 0, 0, 0, 0}ζk
which can be equivalently expressed by utilizing ζk as follows: i=1
  ≤ 0.
E ζkT diag {−1, Θns ,i , 0, 0, 0, 0, 0, 0} ζk ≤ 0, (57)

Authorized licensed use limited to: PAKISTAN INST OF ENGINEERING AND APPLIED SCIENCES. Downloaded on October 08,2022 at 17:10:56 UTC from IEEE Xplore. Restrictions apply.
 476 IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS, VOL. 7, 2021

We now acquire from Lemma 1 (S-procedure) that [20] C. Li, H. Dong, J. Li, and F. Wang, “Distributed kalman filtering for
sensor network with balanced topology,” Syst. Control Lett., vol. 131, 2019,
E{x̃T −1
i,k+1 Pk+1 x̃i,k+1 }
Art. no. 104500.
[21] H. Liu, Z. Wang, W. Fei, and J. Li, “H∞ and l2 − l∞ state estimation for
−ζkT diag{1, 0, 0, 0, 0, 0, 0, 0}ζk ≤ 0, (58) discrete-time delayed memristive neural networks on finite horizon: The
round-robin protocol,” Neural Netw., vol. 132, pp. 121–130, 2020.
[22] L. Liu, L. Ma, J. Zhang, and Y. Bo, “Distributed non-fragile set-
or equivalently, membership filtering for nonlinear systems under fading channels and
bias injection attacks,” Int. J. Syst. Sci., vol. 52, no. 6, pp. 1192–1205,
E{x̃T −1
i,k+1 Pk+1 x̃i,k+1 } ≤ 1, (59) 2021.
[23] S. Liu, Z. Wang, G. Wei, and M. Li, “Distributed set-membership filter-
which indicates the proof is complete. ing for multi-rate systems under the round-robin scheduling over sensor
networks,” IEEE Trans. Cybern., vol. 50, no. 5, pp. 1910–1920, May 2020.
[24] L. Ma, Z. Wang, Y. Chen, and X. Yi, “Probability-guaranteed dis-
REFERENCES tributed filtering for nonlinear systems with innovation constraints
over sensor networks,” IEEE Control Netw. Syst., to be published,
[1] D. P. Bertsekas and I. B. Rhodes, “Recursive state estimation for a doi: 10.1109/TCNS.2021.3049361.
set-membership description of uncertainty,” IEEE Trans. Autom. Control, [25] L. Ma, Z. Wang, H.-K. Lam, and N. Kyriakoulis, “Distributed event-based
vol. AC-16, no. 2, pp. 117–128, Apr. 1971. set-membership filtering for a class of nonlinear systems with sensor
[2] S. Boyd, L. Ghaoui, E. Feron, and V. Balakrishnan, Linear Matrix Inequal- saturations over sensor networks,” IEEE Trans. Cybern., vol. 47, no. 11,
ities in System and Control Theory. Philadelphia, PA, USA: SIAM Studies pp. 3772–3783, Nov. 2017.
in Applied Mathematics, 1994. [26] L. Ma, Z. Wang, Q.-L. Han and Y. Liu, “Distributed filtering for nonlinear
[3] G. Calariore, “Reliable localization using set-value nonlinear filters,” time-delay systems over sensor networks subject to multiplicative link
IEEE Trans. Syst., Man Cybern.-Part A: Systems Humans, vol. 35, no. 2, noises and switching topology,” Int. J. Robust Nonlinear Control, vol. 29,
pp. 189–197, Mar. 2005. pp. 2941–2959, 2019.
[4] R. Carli, A. Chiuso, L. Schenato, and S. Zampieri, “Distributed Kalman [27] L. Ma, Z. Wang, J. Hu, and Q.-L. Han, “Probability-guaranteed envelope-
filtering based on consensus strategies,” IEEE Trans. Sel. Areas Commun., constrained filtering for nonlinear systems subject to measurement out-
vol. 26, no. 4, pp. 622–633, 2010. liers,” IEEE Trans. Autom. Control, vol. 66, no. 7, pp. 3274-3281, Jul.
[5] F. S. Cattivelli and A. H. Sayed, “Diffusion strategies for distributed 2021.
Kalman filtering and smoothing,” IEEE Trans. Autom. Control, vol. 55, [28] J. Mao, Y. Sun, X. Yi, H. Liu, and D. Ding, “Recursive filtering of
no. 9, pp. 2069–2084, Sep. 2010. networked nonlinear systems: A survey,” Int. J. Syst. Sci., vol. 52, no. 6,
[6] Y. Chen, Z. Chen, Z. Chen, and A. Xue, “Observer-based passive control pp. 1110–1128, 2021.
of non-homogeneous Markov jump systems with random communication [29] A. De Paola, S. Gaglio, G. L. Re, F. Milazzo, and M. Ortolani, “Adaptive
delays,” Int. J. Syst. Sci., vol. 51, no. 6, pp. 1133–1147, 2020. distributed outlier detection for WSNs,” IEEE Trans. Cybern., vol. 45,
[7] Y. Chen, Z. Wang, Y. Yuan, and P. Date, “Distributed H∞ filtering for no. 5, pp. 888–899, May 2015.
switched stochastic delayed systems over sensor networks with fading [30] A. Ribeiro, I. Schizas, S. Roumeliotis, and G. Giannakis, “Kalman filtering
measurements,” IEEE Trans. Cybern., vol. 50, no. 1, pp. 2–14, Jan. 2020. in wireless sensor networks,” IEEE Control Syst. Mag., vol. 30, pp. 66–86,
[8] L. El Ghaoui and G. Calafiore, “Robust filtering for discrete-time systems Apr. 2010.
with bounded noise and parametric uncertainty,” IEEE Trans. Autom. [31] S. Seifzadeh, B. Khaleghi, and F. Karray, “Distributed soft-data-
Control, vol. 46, no. 7, pp. 1084–1089, Jul. 2001. constrained multi-model particle filter,” IEEE Trans. Cybern., vol. 45,
[9] G. Franze, F. Tedesco, and W. Lucia, “Resilient control for cyber-physical no. 3, pp. 384–394, Mar. 2015.
systems subject to replay attacks,” IEEE Contr. Syst. Lett., vol. 3, no. 4, [32] B. Shen, Z. Wang, D. Wang, and H. Liu, “Distributed state-saturated
pp. 984–985, Oct. 2019. recursive filtering over sensor networks under round-robin protocol,” IEEE
[10] X. Ge, Q.-L. Han, X. M. Zhang, L. Ding, and F. Yang, “Distributed Trans. Cybern., vol. 50, no. 8, pp. 3605–3615, Aug. 2020.
event-triggered estimation over sensor networks: A survey,” IEEE Trans. [33] B. Shen, Z. Wang, and Y. S. Hung, “Distributed H∞ -consensus filtering in
Cybern., vol. 50, no. 3, pp. 1306–1320, Mar. 2020. sensor networks with multiple missing measurements: The finite-horizon
[11] X. Ge, Q.-L. Han, L. Ding, Y. Wang, and X.-M. Zhang, “Dynamic event- case,” Automatica, vol. 46, no. 10, pp. 1682–1688, Oct. 2011.
triggered distributed coordination control and its applications: A survey [34] Y. Sun, J. Mao, H. Liu, and D. Ding, “Distributed recursive filtering for
of trends and techniques,” IEEE Trans. Syst., Man, Cybern. Syst., vol. 50, discrete time-delayed stochastic nonlinear systems based on fuzzy rules,”
no. 9, pp. 3112–3125, Sep. 2020. Neurocomputing, vol. 400, pp. 412–419, 2020.
[12] X. Ge, Q.-L. Han, M. Zhong, X.-M. Zhang, “Distributed krein space-based [35] H. Tan, B. Shen, K. Peng, and H. Liu, “Robust recursive filtering for
attack detection over sensor networks under deception attacks,” Automat- uncertain stochastic systems with amplify-and-forward relays,” Int. J. Syst.
ica, vol. 109, Nov. 2019, Art. no. 108557 Sci., vol. 51, no. 7, pp. 1188–1199, 2020.
[13] F. Han, Z. Wang, H. Dong, and H. Liu, “Partial-nodes-based scalable H∞ - [36] E. Tian, Z. Wang, L. Zou, and D. Yue, “Probabilistic-constrained fil-
consensus filtering with censored measurements over sensor networks,” tering for a class of nonlinear systems with improved static event-
IEEE Trans. Syst., Man, Cybern. Syst., vol. 51, no. 3, pp. 1892–1903, triggered communication,” Int. J. Robust Nonlinear Control, vol. 29, no. 5,
Mar. 2021. pp. 1484–1498, 2019.
[14] J. Hu, Z. Wang, G.-P. Liu, H. Zhang, and R. Navaratne, “A prediction- [37] V. Ugrinovskii, “Distributed robust filtering with H∞ consensus of esti-
based approach to distributed filtering with missing measurements and mates,” Automatica, vol. 47, no. 1, pp. 1–13, 2011.
communication delays through sensor networks,” IEEE Trans. Syst., Man, [38] M. Wang, L. Huang, and C. Yang, “NN-based adaptive tracking control
Cybern. Syst., to be published, doi: 10.1109/TSMC.2020.2966977. of discrete-time nonlinear systems with actuator saturation and event-
[15] J. Hu, H. Zhang, H. Liu, and X. Yu, “A survey on sliding mode control for triggering protocol,” IEEE Trans. Syst., Man, Cybernetics: Syst., to be
networked control systems,” Int. J. Syst. Sci., vol. 52, no. 6, pp. 1129–1147, published, doi: 10.1109/TSMC.2020.2981954.
2021. [39] G. Wei, W. Li, D. Ding, and Y. Liu, “Stability analysis of covariance
[16] X. Li, F. Han, N. Hou, H. Dong, and H. Liu, “Set-membership filtering for intersection-based Kalman consensus filtering for time-varying systems,”
piecewise linear systems with censored measurements under round-robin IEEE Trans. Syst., Man, Cybern. Syst., vol. 50, no. 11, pp. 4611–4622,
protocol,” Int. J. Syst. Sci., vol. 51, no. 9, pp. 1578–1588, 2020. Nov. 2020.
[17] M. Jorjani, H. Seifi and A. Varjani, “A graph theory-based approach to [40] I. Yaesha, S. Boyarskia, and U. Shaked, “Probability-guaranteed robust
detect false data injection attacks in power system AC state eestimation,” H∞ performance analysis and state-feedback design,” Syst. Control Lett.,
IEEE Trans. Ind. Informat., vol. 17, no. 4, pp. 2465–2475, Apr. 2021. vol. 48, pp. 351–364, 2003.
[18] Q. Li, B. Shen, Z. Wang, and W. Sheng, “Recursive distributed filter- [41] N. D. Zade, S. N. Deshpande, and D. Sita, “Investigative analysis of
ing over sensor networks on Gilbert–Elliott channels: A dynamic event- suboptimal filter for state estimation in object tracking wireless sensor
triggered approach,” Automatica, vol. 113, 2020, Art. no. 108681. network,” IEEE Sens. Lett., vol. 4, no. 10, pp. 2475–1472, Oct. 2020.
[19] X. Li, Y. Zhao, S. Zhang, and X. Fan, “Adaptive particle filter for non- [42] C. Zhao, J. Lam, and H. Lin, “State estimation of CPSs with deception at-
parametric estimation with measurement uncertainty in wireless sensor tacks: Stability analysis and approximate computation,” Neurocomputing,
networks,” Sensors, vol. 16, no. 6, 2016, Art. no. 786. vol. 423, pp. 318–326, 2021.

Authorized licensed use limited to: PAKISTAN INST OF ENGINEERING AND APPLIED SCIENCES. Downloaded on October 08,2022 at 17:10:56 UTC from IEEE Xplore. Restrictions apply.
 MA et al.: PROBABILITY-GUARANTEED DISTRIBUTED SECURE ESTIMATION FOR NONLINEAR SYSTEMS OVER SENSOR NETWORKS 477

[43] D. Zhao, Z. Wang, D. W. C. Ho, and G. Wei, “Observer-based PID security Yun Chen was born in Zhejiang Province, China. He
control for discrete time-delay systems under cyber-attacks,” IEEE Trans. received the B.E. degree in thermal engineering from
Syst., Man, Cybern. Syst., vol. 51, no. 6, pp. 3926–3938, Jun. 2021. the Central South University of Technology (Central
[44] D. Zhao, Z. Wang, G. Wei, and Q.-L. Han, “A dynamic event-triggered South University), Changsha, China, in 1999, and the
approach to observer-based PID security control subject to deception M.E. degree in engineering thermal physics and the
attacks,” Automatica, vol. 120, 2020, Art. no. 109128. Ph.D. degree in control science and engineering from
[45] Z. Zhao, Z. Wang, L. Zou, and J. Guo, “Set-membership filtering for Zhejiang University, Hangzhou, China, in 2002 and
time-varying complex networks with uniform quantisations over ran- 2008, respectively.
domly delayed redundant channels,” Int. J. Syst. Sci., vol. 51, no. 16, From August 2009 to August 2010, he was a Visit-
pp. 3364–3377, 2020. ing Fellow with the School of Computing, Engineer-
[46] K. Zhu, J. Hu, Y. Liu, N. D. Alotaibi, and F. E. Alsaadi, “On 2 -∞ ing and Mathematics, University of Western Sydney,
output-feedback control scheduled by stochastic communication protocol Australia. From December 2016 to December 2017, he was an Academic Visitor
for two-dimensional switched systems,” Int. J. Syst. Sci., to be published, with the Department of Mathematics, Brunel University London, U.K. In 2002,
doi: 10.1080/00207721.2021.1914768. he joined Hangzhou Dianzi University, China, where he is currently a Professor.
[47] L. Zou, Z. Wang, H. Geng, and X. Liu, “Set-membership filtering subject His research interests include stochastic and hybrid systems, robust control, and
to impulsive measurement outliers: A recursive algorithm,” IEEE/CAA J. filtering.
Automatica Sinica, vol. 8, no. 2, pp. 377–388, Feb. 2021.
[48] L. Zou, Z. Wang, J. Hu, Y. Liu, and X. Liu, “Communication-
protocol-based analysis and synthesis of networked systems: Progress,
prospects and challenges,” Int. J. Syst. Sci., to be published,
doi: 10.1080/00207721.2021.1917721.
Xiaojian Yi was born in 1987. He received the B.S.
degree in control technology from the North Univer-
Lifeng Ma received the B.Sc. degree in automation sity of China, Taiyuan, China, in 2010, and the M.S.
and Ph.D. degrees in reliability engineering from the
from Jiangsu University, Zhenjiang, China, in 2004
Beijing Institute of Technology, Beijing, China, in
and the Ph.D. degree in control science and engi-
2012 and 2016, respectively. During 2015–2016, he
neering from the Nanjing University of Science and
was a jointly trained Ph.D. student with the Univer-
Technology, Nanjing, China, in 2010. From August
sity of Ottawa, Ottawa, ON, Canada, to study robot
2008 to February 2009, he was a Visiting Ph.D. Stu-
reliability and maintenance. From 2016 to 2020, he
dent with the Department of Information Systems and
Computing, Brunel University London, U.K. From was an Associate Professor with China North Vehicle
Research Institute. He is currently an Associate Pro-
January 2010 to April 2010 and May 2011 to Septem-
fessor with the Beijing Institute of Technology. He is the author of two books
ber 2011, he was a Research Associate with the
and more than 100 articles, and is also the holder of eight patents. His research
Department of Mechanical Engineering, University
of Hong Kong, Hong Kong. From March 2015 to February 2017, he was a interests include system reliability analysis, intelligent control, fault diagnosis,
and health management.
Visiting Research Fellow with King’s College London, U.K.
He is currently a Professor with the School of Automation, Nanjing University
of Science and Technology. He has authored or coauthored more than 50 papers
in refereed international journals. His current research interests include control
and signal processing, machine learning, and deep learning. He is the Editor of
the Neurocomputing and International Journal of Systems Science.

Zidong Wang (Fellow, IEEE) was born in Jiangsu,

China, in 1966. He received the B.Sc. degree in math-
ematics from Suzhou University, Suzhou, China, in
1986, and the M.Sc. degree in applied mathematics
and the Ph.D. degree in electrical engineering from
the Nanjing University of Science and Technology,
Nanjing, China, in 1990 and 1994, respectively.
He is currently a Professor of dynamical systems
and computing with the Department of Computer
Science, Brunel University London, U.K. From 1990
to 2002, he held teaching and research appointments
with universities in China, Germany and the U.K. He has authored or coauthored
more than 600 papers in international journals. His research interests include
dynamical systems, signal processing, bioinformatics, control theory and appli-
cations. He is a holder of the Alexander von Humboldt Research Fellowship
of Germany, the JSPS Research Fellowship of Japan, William Mong Visiting
Research Fellowship of Hong Kong.
Prof. Wang serves (or has served) as the Editor-in-Chief of the International
Journal of Systems Science, Neurocomputing, Systems Science & Control Engi-
neering, and an Associate Editor for 12 international journals, including the IEEE
TRANSACTIONS ON AUTOMATIC CONTROL, IEEE TRANSACTIONS ON CONTROL
SYSTEMS TECHNOLOGY, IEEE TRANSACTIONS ON NEURAL NETWORKS, IEEE
TRANSACTIONS ON SIGNAL PROCESSING, and IEEE TRANSACTIONS ON SYS-
TEMS, MAN, AND CYBERNETICS-PART C. He is a Member of the Academia
Europaea, a Member of the European Academy of Sciences and Arts, an Aca-
demician of the International Academy for Systems and Cybernetic Sciences,
a Fellow of the Royal Statistical Society, and a Member of program committee
for many international conferences.

Authorized licensed use limited to: PAKISTAN INST OF ENGINEERING AND APPLIED SCIENCES. Downloaded on October 08,2022 at 17:10:56 UTC from IEEE Xplore. Restrictions apply.