IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL.

19, 2024 1883

Detection and Mitigation of Position Spoofing

Attacks on Cooperative UAV Swarm Formations
Siguo Bi , Kai Li , Senior Member, IEEE, Shuyan Hu , Member, IEEE, Wei Ni , Fellow, IEEE,
Cong Wang, and Xin Wang , Fellow, IEEE

Abstract— Detecting spoofing attacks on the positions of

unmanned aerial vehicles (UAVs) within a swarm is challeng-
ing. Traditional methods relying solely on individually reported
positions and pairwise distance measurements are ineffective
in identifying the misbehavior of malicious UAVs. This paper
presents a novel systematic structure designed to detect and mit-
igate spoofing attacks in UAV swarms. We formulate the problem
of detecting malicious UAVs as a localization feasibility problem,
leveraging the reported positions and distance measurements.
To address this problem, we develop a semidefinite relaxation
(SDR) approach, which reformulates the non-convex localization
problem into a convex and tractable semidefinite program (SDP).
Additionally, we propose two innovative algorithms that leverage
the proximity of neighboring UAVs to identify malicious UAVs
effectively. Simulations demonstrate the superior performance Fig. 1. An illustration of the attack model, where a malicious UAV falsifies
its position and broadcasts the fake position to the other benign UAVs.
of our proposed approaches compared to existing benchmarks.
Our methods exhibit robustness across various swarm networks,
showcasing their effectiveness in detecting and mitigating spoof-
ing attacks. Specifically, the detection success rate is improved UAVs, and their flexibly adjustable positions conducive to
by up to 65%, 55%, and 51% against distributed, collusion, and line-of-sight (LOS) communications, facilitated by rapid tech-
mixed attacks, respectively, compared to the benchmarks. nological advancements [6]. The varieties of practical needs
Index Terms— Malicious UAV detection, position spoofing for UAV swarms further ignite and necessitate the protection
attack, cooperative localization, semidefinite programming. of security for UAV swarms [7]. For instance, the reliability
of information propagation has been analyzed in large-scale
I. I NTRODUCTION
networks, including UAV swarms [8]. The connectivity of a
R ECENTLY, there has been a widespread utilization of
unmanned aerial vehicles (UAVs) [1], including parcel
delivery [2], radio surveillance [3], [4], and rescue mis-
UAV swarm has been studied in the presence of jamming
attacks from the ground [9]. To enhance reliability and mitigate
potential flight collisions, it is crucial to establish a formation
sions [5]. This is due to the affordability and endurance of
flight and coordination among UAVs [10]. In the formation
Manuscript received 17 July 2023; revised 18 October 2023; flight of a UAV swarm, individual UAVs rely on position
accepted 3 December 2023. Date of publication 8 December 2023; reports from their peers and their pairwise distance measure-
date of current version 26 December 2023. This work was supported in part
by the National Natural Science Foundation of China under Grant 62231010, ments with neighboring UAVs to maintain inter-UAV distances
Grant 62071126, and Grant 62101135; and in part by the Innovation Program and avoid collisions. Compromised or malicious UAVs can
of Shanghai Municipal Science and Technology Commission under Grant launch position spoofing attacks, potentially leading to catas-
21XD1400300. The work of Kai Li was supported by the Real-Time and
Embedded Computing Systems Research Centre (CISTER) Research Unit trophic consequences for the UAV swarm [11]. A malicious
under Grant UIDP/UIDB/04234/2020 and in part by Project ADANET UAV might transmit a deceptive position report, misleading
financed by National Funds through Portuguese Foundation for Science other UAVs while simultaneously concealing its true location,
and Technology [FCT/Portuguese Ministry of Science, Technology, and
Higher Education (MCTES)] under Grant PTDC/EEICOM/3362/2021. The as illustrated in Fig. 1. Such conditions can disrupt the
associate editor coordinating the review of this manuscript and approving it control mechanism that maintains swarm formation, resulting
for publication was Prof. Haijun Zhang. (Corresponding author: Xin Wang.) in disorders [12].
Siguo Bi, Shuyan Hu, Cong Wang, and Xin Wang are with
Fudan University, Shanghai 200433, China (e-mail: [email protected]; Detecting and identifying malicious UAVs within a swarm is
[email protected]; [email protected]; [email protected]). challenging. This problem seeks to establish whether a feasible
Kai Li is with the Division of Electrical Engineering, Department of position realization for each UAV that aligns with all reported
Engineering, University of Cambridge, CB3 0FA Cambridge, U.K., and also
with the Real-Time and Embedded Computing Systems Research Centre distances and measurements exists. Such a feasibility problem
(CISTER), Porto 4249-015, Portugal (e-mail: [email protected]). is non-trivial and non-convex, and has never been studied
Wei Ni is with the Commonwealth Scientific and Industrial and addressed in the existing literature. As delineated in this
Research Organization (CSIRO), Sydney, NSW 2122, Australia (e-mail:
[email protected]). paper, the problem can be transformed into a convex semidef-
Digital Object Identifier 10.1109/TIFS.2023.3341398 inite program (SDP), allowing for efficient use of convex
1556-6021 © 2023 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: XIDIAN UNIVERSITY. Downloaded on December 09,2024 at 13:00:07 UTC from IEEE Xplore. Restrictions apply.
1884 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 19, 2024

optimization solvers in polynomial time. However, solving TABLE I

the SDP problem alone does not enable the identification of N OTATION AND D EFINITION
individual malicious UAVs. On the one hand, the number
of malicious UAVs is typically unknown and needs to be
detected. On the other hand, the effectiveness of SDP can
be penalized by the interdependence among the positions of
neighboring UAVs.
To address these challenges and precisely identify mali-
cious UAVs, this paper proposes two new algorithms: the
Cooperative Detection and Identification (CDI) algorithm and
the Enhanced CDI (E-CDI) algorithm. The CDI algorithm
initiates its process by creating sets of possible malicious
and benign UAVs. Subsequently, it combines the selected
potentially malicious UAVs with the benign set, establishing a
connected sub-network for the SDP-based position feasibility
check. If all the neighboring UAVs of a selected UAV are
themselves malicious, the CDI algorithm may misjudge the
UAV as malicious, as attempting to localize a sub-network
with an entire malicious neighborhood is inherently unfeasi-
ble. In contrast, the E-CDI algorithm conducts an additional
localization feasibility check on each individual UAV in the
neighborhood, compared to the CDI algorithm. By this means,
collusion attacks launched by multiple closely located, mali-
cious UAVs can be detected and mitigated.
Compared to the existing relevant works, e.g., [11], [12],
[13], and [14], the new contributions of this paper include: The rest of this paper is organized as follows. Section II
1) To detect position spoofing attacks, we propose a novel reviews the related works. Section III formulates and con-
mechanism for malicious UAV detection and identifi- vexifies the malicious UAV’s misbehavior detection problem.
cation, where we cast the challenging malicious UAV In Section IV, two efficient iterative algorithms are proposed to
detection problem as a localization feasibility problem. identify malicious UAVs. Section V provides numerical results
2) A semidefinite relaxation (SDR) approach is put forth to evaluate the proposed algorithm, followed by conclusions
to transform the non-convex feasibility problem into in Section VI.
a convex problem. The presence of malicious UAVs Notation: Upper- and lower-case boldface symbols denote
can then be efficiently ascertained by evaluating the matrices and vectors, respectively; | · | takes the absolute value
feasibility of the convex problem. if a scalar is concerned or the cardinality if a set is concerned;
3) We develop two iterative algorithms, i.e., CDI and ∥·∥ denotes ℓ2 -norm; (·)
ˆ indicates a reported, noise-corrupted
E-CDI, to identify malicious UAVs by leveraging the version of (·). The notation used is collated in Tab. I.
proximity of neighboring UAVs.
• The CDI algorithm dynamically merges selected II. R ELATED W ORK
potentially malicious UAVs into the benign set to A. Spoofing Attacks on UAVs
form a connected positioning sub-network. This
Spoofing attacks on UAVs have been extensively inves-
sub-network is used to determine whether the
tigated in the recent literature. However, most works have
selected UAV is malicious.
focused on the direct hijack of the Global Positioning System
• The E-CDI algorithm enhances identification effi-
(GPS) of a specific single UAV. Aiming at identifying fake
ciency by further assessing each neighboring UAV
GPS coordinates due to the hijack of the GPS communication
in the neighborhood of a potentially malicious UAV.
software, the authors of [15] proposed a convolutional neural
As a result, collusion attacks launched by multiple
network (CNN) integrated with a recurrent neural network
closely located, malicious UAVs can be detected.
(RNN) to predict a vehicle’s real-time trajectory based on
Both algorithms are designed to conclude within a finite the data from multiple sensors. With a similar purpose of
number of iterations and exhibit robust performance handling GPS spoofing attacks, the authors of [16] proposed
across various network configurations of UAV swarms. a two-step approach based on data sensed and fused from
Extensive simulations demonstrate that the proposed CDI and distributed radar ground stations equipped with a local tracker.
E-CDI algorithms achieve superior performance on classic The approach consists of spoofing detection and mitigation.
metrics compared to the benchmark techniques. Under the pro- In the spoofing detection step, a track-to-track association
posed algorithms, the detection success rate can be improved approach was adopted to detect spoofing attacks with fused
by up to 65%, 55%, and 51% against distributed, collusion, data from UAVs and a local tracker. In the mitigation step, the
and mixed attacks, respectively, compared to their benchmarks. fused data was input to a controller to mitigate the spoofing

Authorized licensed use limited to: XIDIAN UNIVERSITY. Downloaded on December 09,2024 at 13:00:07 UTC from IEEE Xplore. Restrictions apply.
BI et al.: DETECTION AND MITIGATION OF POSITION SPOOFING ATTACKS 1885

attack detected. The proposed two-step approach was reported

to achieve almost the same accuracy as GPS efficiently.
To enhance the reliability of flight controllers when the
UAV is under GPS spoofing attack, the authors of [17]
utilized an extended Kalman filter (EKF)-based approach.
They investigated the impact of GPS spoofing on the EKF
estimation and the UAV itself under different levels of attack
strength. It was reported that the classic EKF-based approach
can tolerate small errors from spoofing attacks, but can be
inefficient when the attack intensifies. Similar works on GPS-
related spoofing attacks on a specific single UAV can be found
in [13], [18], [19], and [20], and spoofing attacks related to the Fig. 2. An illustration of the collusion spoofing attack model, where two
time-of-arrival (TOA) or time difference-of-arrival (TDOA) malicious UAVs falsify their positions to be within the one-hop neighborhood
can be found in [21]. of the benign UAV.
The security issue of UAV swarms has attracted increasing
attention. In order to mitigate the navigation spoofing attacks
on aerial formations, the authors of [22] proposed a cascaded as a localization feasibility problem. By applying the SDR,
estimation algorithm used for concurrent GPS spoofing detec- we convexify the feasibility problem into a convex problem.
tion and localization. An attack detection module was based on
the consistency of multimodal measurement to realize thresh- A. Threat Model
old tests. A localization module was then used for a decision Consider a swarm of UAVs executing a routine cruising
based on remarkable differences between safe and under-attack mission, during which the UAVs cooperatively maintain a
conditions of UAV self-localization. The cascaded approach specific formation to prevent collisions. Each UAV within the
can achieve a safe self-localization for a UAV swarm under a swarm communicates its position, ascertained by the GPS,
spoofing attack. Aiming at solving the GPS spoofing attack in to its counterparts. Each UAV also conducts relative distance
a UAV swarm, the authors of [23] proposed a security-aware measurements with the other UAVs that are within the per-
monitoring method to monitor the potential malicious UAVs missible communication range of the UAV. Random receiver
and protect the benign ones from attacks. The method was noises or GPS errors can corrupt these distance measurements,
implemented by the received-signal-strength-indicator (RSSI)- rendering the reported positions inaccurate. During the forma-
based triangulation. tion flight, each UAV adjusts its flight position based on the
reported positions and distance measurements of neighboring
B. Cooperative Network Localization
UAVs, thereby averting potential flight collisions.
Position-related spoofing attacks destroy the localization of Malicious UAVs, under the control of attackers, have the
UAVs in a UAV swarm, since a UAV swarm can be considered capability to fabricate their position information and dissemi-
a cooperative network. SDP, an efficient convex optimization nate this information among all other UAVs. More precisely,
approach, has been extensively applied to cooperative network malicious UAVs can initiate a spoofing attack by falsely report-
localization. Employing the SDP, the authors of [24] proposed ing their positions within the detection measurement range of
a novel difference-of-convex (DC)-based algorithm to achieve benign UAVs. This deliberate misrepresentation of positions
accurate cooperative localization. The authors of [25] proposed can directly disrupt the formation. Moreover, malicious UAVs
an SDP-based method to estimate the relative transformation may target a specific benign UAV by deceptively reporting
of a robot in a cooperative robotic swarm. The SDP-based their positions within the detection measurement range of that
method could achieve global optimality and scalability. The UAV, which is a tactic known as a collusion attack; see Fig. 2.
authors of [26] developed an efficient SDP-based scheduling This coordinated misrepresentation is aimed at framing a target
strategy to optimize UAV deployment in intelligent trans- UAV. Given the substantial evidence presented through this
portation cooperative networks. More SDP-based cooperative deceptive conspiracy, the swarm may erroneously conclude
localization techniques can be found in [14], [27], and [28]. that the framed UAV is perpetrating a spoofing attack.
Unlike existing works that rely heavily on extensive training The two above-mentioned attacks, i.e., the distributed attack
using historical data, we put forth an SDP-based UAV misbe- and the collusion attack, can be amalgamated to initiate a
havior detection mechanism with no need for historical data. mixed spoofing attack; see Fig. 3. In this composite attack,
The proposed SDP-based mechanism detects and identifies some malicious UAVs execute a distributed attack while the
malicious UAVs that misreport their positions by leveraging remaining UAVs engage in a collusion attack. The mixed
the proximity of neighboring UAVs. This mechanism is appli- nature of this attack significantly intensifies its severity, poten-
cable regardless of the specific type of localization signals tially hastening the breakdown of the entire formation.
hijacked and spoofed, including GPS, TDA, or TDOA.

III. S YSTEM M ODEL AND P ROBLEM F ORMULATION B. Problem Statement

In this section, we first provide the threat model considered We propose to formulate the problem of identifying mali-
and formally formulate the malicious UAV detection problem cious UAVs as a localization feasibility problem. If localization

Authorized licensed use limited to: XIDIAN UNIVERSITY. Downloaded on December 09,2024 at 13:00:07 UTC from IEEE Xplore. Restrictions apply.
1886 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 19, 2024

measurement error. Nonetheless, the right-hand side (RHS) of

constraint (1c), i.e., ( d2 )2 , can adapt to the needs of different
measurement devices.

C. Proposed SDP-Based Reformulation

Because of the non-convex constraints (1a)–(1c), finding X
in (1) is a non-convex feasibility checking problem, which is
difficult to solve. To convexify (1a)–(1c), auxiliary variables,
denoted by αi j , ∀i, j ∈ [1, N ], are introduced to substitute
those non-convex parts in (1a)–(1c). As a result, the non-
convex feasibility problem in (1) can be rewritten as
Fig. 3. An illustration of the mixed spoofing attack model, where two
malicious UAVs conduct a collusion attack while the other malicious UAVs
falsify their positions.
findX
s.t. αi j < d 2 , ∀i ∈ N, j ∈ Ni , (2a)
αii ≤ ϵ, ∀i ∈ N, (2b)
is infeasible under the positions and distance measurements d
reported, there is at least one malicious UAV attacking |r̂i2j − αi j | < ( )2 , ∀i ∈ N, j ∈ Ni , (2c)
2
the UAV swarm in an attempt to compromise the swarm
formation. ||x i − x̂ j ||2 = αi j , ∀i ∈ N, j ∈ Ni ∪ {i}, (2d)
Let X := {x 1 , . . . , x N } define the three-dimensional (3D) where constraints (2a) and (2d) are homogenized from (1a),
coordinates of the UAVs in an N -UAV swarm studied, where and (2b) is homogenized from (1b). Constraints (2a)–(2c)
x i ∈ R3×1 is the unknown actual 3D coordinates of UAV i. are affine and convex. Constraint (2d) is still non-convex.
Let N = {1, . . . , N } collect the indexes of the N UAVs. To convexify (2d), we rewrite ||x i − x̂ j ||2 in a matrix form as
The permissible range for pairwise distance measurement is   
T I3 X x̂ j
denoted by d. The reported position of UAV i is given as xˆi , 2 T
∥x i − x̂ j ∥ = [ x̂ j − ei ] , (3)
which can be contaminated by the measurement noise, i.e., XT Y −ei
wi ∼ N (0, σi2 I 3 ). x̂ i = x i +wi , ∀i. Here, N (0, σi2 I 3 ) stands where ei ∈ R N ×1 is a vector whose i-th element is “1” and
for the zero-mean Gaussian distribution with the variance of the rest are all “0”. X ∈ R3×N is a 3 × N matrix with its i-th
σi2 I 3 and I 3 is the 3 × 3 identity matrix. column being x i . Moreover,
The feasibility problem of finding a solution of X can be
formulated as Y = X T X ∈ R N ×N . (4)

findX As a result, finding X in problem (2) can be equivalently

reformulated as
s.t. ∥x i − x̂ j ∥2 < d 2 , ∀i ∈ N, j ∈ Ni , (1a)
|x i − x̂ i ∥ ≤ ϵ, ∀i ∈ N,
2
(1b) find X, Y
d s.t. αi j < d 2 , ∀i ∈ N, j ∈ Ni , (5a)
|r̂i2j − ∥x i − x̂ j ∥2 | < ( )2 , ∀i ∈ N, j ∈ Ni , (1c)
2 αii ≤ ϵ, ∀i ∈ N, (5b)
where Ni collects all one-hop neighbors of UAV i and can d
|r̂i2j − αi j | < ( )2 , ∀i ∈ N, j ∈ Ni , (5c)
be obtained based on the reports of UAV i; ϵ ≥ 0 is a  2  
small constant; and r̂i j denotes the reported pairwise distance I3 X x̂ j
[ x̂ Tj − eiT ] = αi j , ∀i ∈ N, j ∈ Ni ∪ {i},
between UAVs i and j contaminated with measurement noise, X T Y −ei
e.g., wi j ∼ N (0, σi2j ). Here, σi2j is the variance of the distance (5d)
measurement error. T
Constraint (1a) defines that the relative distance measure- Y = X X. (5e)
ment between UAV i and the reported position of UAV j, i.e., Here, constraints (5d) and (5e) are non-convex. Yet, constraint
∥x i − x̂ j ∥, has to be within the permissible communication (5e) can be relaxed as [24] and [29]
range, d, if UAV i is a one-hop neighbor of UAV j and
can hear its ranging signals. Constraint (1b) dictates that the Y ⪰ X T X, (6)
difference between x i and xˆi is smaller than a pre-specified where “⪰” stands for element-wise inequality. According to
threshold ϵ, ensuring that the model must rely on the individual Schur complement [30], (6) is equivalent to
reported position to output a solution (if such a solution does  
not exist, it is reasonable to suspect there exist misreported I3 X
⪰ 0, (7)
UAV positions). Constraint (1c) indicates that the difference XT Y
between the reported pairwise distance and the Euclidean Further let Z denote the left-hand side (LHS) of (7), yielding
distance (between the estimated and reported positions) should  
be smaller than half of the distance measurement range, which I3 X
Z= ⪰ 0. (8)
is also considered to be the maximum tolerable distance XT Y

Authorized licensed use limited to: XIDIAN UNIVERSITY. Downloaded on December 09,2024 at 13:00:07 UTC from IEEE Xplore. Restrictions apply.
BI et al.: DETECTION AND MITIGATION OF POSITION SPOOFING ATTACKS 1887

The relaxation of (5e) to (6) is tight if Rank(Z) = 3. Algorithm 1 The Proposed CDI Algorithm
Also, define Input: B; M; xˆi , ∀i ∈ [1, N ]; rˆi j ,
∀i, j ∈ [1, N ], i ̸ = j; d; ϵ.
 
x̂ j
Ĝ i j = [ x̂ Tj − eiT ]. (9)
−ei Output: M
1 Set k ← 1;
Based on (8) and (9), (5d) can be rewritten as
   2 while M can be further reduced do
I3 X x̂ j
[ x̂ Tj − eiT ] = Tr(Ĝ i j Z) = αi j , (10) 3 Select the k-th UAV of M and the set of its
XT Y −ei one-hop neighbors Nk ;
Based on (6)–(10), the feasibility problem (5) is further 4 Construct T ← B ∪ {k, Nk };
equivalently rewritten as the following feasibility problem: 5 Apply T to (12), and check feasibility using SDP.
find Z 6 if problem (12) is feasible T then
7 M ← M \ {k, Nk };
s.t. Z 1:3,1:3 = I 3 , (11a) 8 B ← B ∪ {k, Nk };
αi j < d ,
2
∀i ∈ N, j ∈ Ni , (11b) 9 end
αii ≤ ϵ, = ∀i ∈ N, (11c) 10 Set k ← (k + 1) mod |M|;
11 end
d
|r̂i2j − αi j | < ( )2 , ∀i ∈ N, j ∈ Ni , (11d)
2
Tr(Ĝ i j Z) = αi j , ∀i ∈ N, j ∈ Ni ∪ {i}, (11e)
A. Initialization of Malicious UAV Set
Z ⪰ 0, (11f)
Let M and B denote the sets of malicious and benign UAVs,
Rank(Z) = 3. (11g)
respectively. M ∪ B = N. Based on E r and E n , we propose
where constraint (11a) enforces the upper left 3 × 3 block to initialize M and B, as follows.
of Z to be an identity matrix, ensuring that the rank of the We come up with two Euclidean matrices, i.e., the generated
solution is at least three. Constraints (11a), (11f), and (11g) Euclidean matrix from individual reported positions, denoted
are equivalently derived from (5e). This is because both (11a) by E r ∈ R N ×N , and the detected Euclidean distances matrix
and (11f) constrain Z to be symmetric and in the form of (8), contaminated with noise, denoted by E n ∈ R N ×N , as
while rank constraint (11g) forces the lower right N × N block 
ρ11 ∥ xˆ1 − xˆ1 ∥ ··· ρ1N ∥ xˆ1 − xˆN ∥

of Z, i.e., Y in (8), to be X T X, according to classic linear .. .. ..
Er =  . . .  , (13)
 
algebra theory.
Dropping the rank constraint (11g), we have the following ρ N 1 ∥ xˆN − xˆ1 ∥ · · · ρ N N ∥ xˆN − xˆN ∥
SDR problem:
 
r̂11 · · · r̂1N
find Z E n =  ... ..
.
..  ,
.  (14)


s.t. Z 1:3,1:3 = I 3 , (12a) r̂ N 1 ··· r̂ N N

αi j < d , ∀i ∈ N, j ∈ Ni ,
2
(12b) where ρi j indicates if UAVs i and j are directly connected.
αii ≤ ϵ, ∀i ∈ N, (12c) ρi j = 1, if UAVs i and j are within each other’s permissible
d communication range, i.e., r̂i j > 0; otherwise, ρi j = 0. In this
|r̂i2j − αi j | < ( )2 , ∀i ∈ N, j ∈ Ni , (12d) sense, E r can be a sparse matrix (like E n ), depending on the
2
communication range of the UAV.
Tr(Ĝ i j Z) = αi j , ∀i ∈ N, j ∈ Ni ∪ {i}, (12e) We can carry out element-wise comparisons between E r
Z ⪰ 0. (12f) and E n . Specifically, if the (i, j)-th elements of the two
Problem (12) is convex and can be efficiently solved using matrices have a smaller difference than the pre-specified
off-the-shelf CVX solvers, e.g., MATLAB CVX toolbox [31]. threshold d2 , i.e., (1c) is unsatisfied, then UAVs i and/or j
Clearly, problem (12) is a relaxed (but generally tight) version are potentially malicious. Both of the UAVs are added into
of the original feasibility problem (1), with a larger feasible M, i.e., M = M ∪ {i, j}. After all N 2 elements of E r and E n
solution region. If the problem in (12) is infeasible, i.e., are assessed, the initial M is obtained. B can be accordingly
no feasible solution exists for the problem in (12), then initialized to be B = N \ M.
problem (1) is surely infeasible. As a result, we can detect
whether at least one malicious UAV misreports its position by B. Proposed CDI Algorithm
checking the feasibility of the problem in (12). As summarized in Algorithm 1, we propose to assess the
potentially malicious UAVs in M one after another and move
IV. P ROPOSED A PPROACH FOR M ALICIOUS UAV those actually benign from M to B until both M and B
I DENTIFICATION stop changing. When assessing a UAV, i.e., UAV k, from M,
Leveraging the SDP problem in (12), we proceed to develop we also consider its one-hop neighbors. Part of Nk may belong
two new algorithms, CDI and E-CDI, to exploit the proxim- to B, and the rest belong to M.
ity of adjacent UAVs to facilitate the effective detection of We apply the feasibility checking problem in (12) to the
malicious UAVs and eliminate spoofing attacks. collection of B, {k}, and Nk , i.e., B∪ {k, Nk }. If the problem is

Authorized licensed use limited to: XIDIAN UNIVERSITY. Downloaded on December 09,2024 at 13:00:07 UTC from IEEE Xplore. Restrictions apply.
1888 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 19, 2024

feasible, UAV k and its one-hop neighbors Nk are benign. They Algorithm 2 The Proposed E-CDI Algorithm
can be removed from M and added to B; i.e., M = M\{k, Nk } Input: B; M; xˆi , ∀i ∈ [1, N ]; rˆi j ,
and B = B∪{k, Nk }. Otherwise, they remain in M. The reason ∀i, j ∈ [1, N ], i ̸ = j; d; ϵ.
for considering a potentially malicious UAV k together with Output: M
its one-hop neighbors Nk is to increase the chance that UAV k 1 Set k ← 1;
is connected to B. Therefore, the feasibility checking problem 2 while M can be further reduced do
can be meaningfully carried out. In the case where UAV k and 3 Select the k-th UAV in M and the set of its
its one-hop neighbors Nk are disconnected from B (in other one-hop neighbors Nk ;
words, Nk and their neighbors all belong to M), then UAV k 4 Construct T ← B ∪ {k, Nk };
and its one-hop neighbors Nk remain in M. 5 Apply T to (12), and check feasibility using SDP.
In this way, we repeatedly assess the remaining UAVs in 6 if problem (12) is feasible upon T then
M until M cannot be further reduced. This algorithm can 7 M ← M \ {k, Nk };
quickly detect and identify malicious UAVs; but may overkill, 8 B ← B ∪ {k, Nk };
i.e., misjudge a benign UAV to be malicious in the situation 9 else
where the benign UAV only has a malicious one-hop neighbor 10 Set T1 ← {k, Nk }; i ← 1;
since they are always assessed together for feasibility and 11 while i ≤ |T1 | do
cannot be individually arbitrated. In this sense, the algorithm 12 Select the i-th UAV of {k, Nk }, denoted
is conservative and can be overprotective. by πi ;
13 Construct T2 ← B ∪ {πi };
C. Proposed E-CDI Algorithm 14 Apply T2 to (12) and check feasibility
A key difference between the E-CDI algorithm and the CDI using SDP;
algorithm (Algorithm 1) is that the E-CDI algorithm assesses 15 if problem (12) is feasible upon T2 then
each of the potentially malicious one-hop neighbors of a UAV 16 M ← M \ {πi };
belonging to M individually, each time the UAV and its one- 17 B ← B ∪ {πi };
hop neighbors fail the feasibility check. Specifically, each of 18 end
UAV k and its potentially malicious one-hop neighbors in Nk 19 i ← i + 1;
are assessed by temporarily joining B for feasibility check 20 end
again. Those that turn out to be benign are removed from 21 end
M and added to B. By this means, each connected malicious 22 Set k ← (k + 1) mod |M|;
UAV can be detected and identified. The details are provided 23 end
in Algorithm 2. The flowchart of the proposed CDI/E-CDI
algorithm is provided in Fig. 4.
Another key difference is that the E-CDI algorithm is able TABLE II
to detect collusion attacks, while the CDI algorithm cannot. S IMULATION PARAMETERS AND C ONFIGURATION
This is because the E-CDI assesses individual UAVs in a
neighborhood {k, Nk } if the neighborhood is detected to be
infected by malicious UAVs in the neighborhood. As a result,
the malicious UAVs (or UAVs that cannot be confirmed benign
due to their poor connectivity to other benign UAVs) can
be individually assessed and verified. In contrast, the CDI
algorithm may not achieve this since its assessment is based
on neighborhoods {k, Nk }, ∀k ∈ M.
d = 0.3. Note that both the reported positions and reported
V. S IMULATION R ESULTS distance measurements are contaminated with additive Gaus-
In this section, we consider three types of spoofing attacks to sian noises, wi ∼ N (0, 10−6 I 3 ) [32] and wi j ∼ N (0, 10−6 ),
gauge the capability of the proposed algorithm to counteract respectively. The key parameters of the simulations are sum-
these attacks. We conduct extensive simulations to compre- marized in Tab. II.
hensively evaluate the proposed algorithms in comparison We assess the performances of the proposed algorithms
with the established benchmarks on the key factors, i.e., the against three types of position spoofing attacks, as follows.
number of malicious UAVs, the scale of the network, distance
• Distributed spoofing attack Under this attack, several
measurement noise, and measurement distance.
malicious UAVs independently misreport their positions
in an attempt to compromise the UAV swarm formation.
A. Simulation Setting • Collusion attack Under this attack, several malicious
We consider a UAV swarm with the UAVs’ positions UAVs conspire to frame some benign UAVs and make
randomly generated according to a uniform distribution inside them falsely identified as malicious. Based on the
a unit cube [−0.5, +0.5]3 . The malicious UAVs are randomly reported positions from targeted benign UAVs, the mali-
chosen from the nodes. The distance measurement range is cious UAVs misreport their positions to be within the

Authorized licensed use limited to: XIDIAN UNIVERSITY. Downloaded on December 09,2024 at 13:00:07 UTC from IEEE Xplore. Restrictions apply.
BI et al.: DETECTION AND MITIGATION OF POSITION SPOOFING ATTACKS 1889

Fig. 4. The flowchart of the proposed CDI/E-CDI algorithm, where the shaded part accounts for the CDI algorithm, which is part of the more comprehensive
E-CDI algorithm.

neighborhood of those benign UAVs, though they can be malicious UAVs and the sampled UAVs as the output of
far away from the UAVs under attack. the approach.
• Mixed spoofing attack Under this attack, malicious The performance metrics considered are Precision, Recall,
UAVs launch attacks in both distributed and collu- and F1. The three classic metrics are given by [34]
sive fashions. Specifically, some of the malicious UAVs
independently carry out distributed spoofing attacks to |Q p ∩ Q t |
compromise the swarm formation. The rest of the mali- Precision = , (15a)
|Q p |
cious UAVs cooperate to further impair or corrode the |Q p ∩ Q t |
integrity of the UAV swarm. Recall = , (15b)
|Q t |
The benchmark algorithms considered are 2 × Precision × Recall
• NLOS-based approach: This approach [33] treats errors F1 = , (15c)
Precision + Recall
induced by the misbehavior of malicious UAVs as a
variant of NLOS, since NLOS and spoofed positions are where Q p stands for the predicted set by a specific algorithm,
alike, i.e., causing considerable deviations from the gen- Q t denotes the ground-truth test set, and |·| denotes cardinality.
uine positions. However, NLOS is a path error involving To evaluate the effect of the network topology, we consider
two UAVs in a swarm. Therefore, we mildly adjust it two other metrics, including “malicious ratio”, i.e., the ratio
to suit comparison by random sampling according to the of the number of malicious UAVs detected initially (as done
scale of the test sample. in Section IV-A) to the total number of UAVs, denoted by
• Random approach: This approach directly relies on R M = |M|/|N|. Correspondingly, the “benign ratio”, i.e., the
the earlier potential candidate set of malicious UAVs ratio of the number of UAVs initially determined as benign to
to conduct random sampling adjusted to the number of the total number of UAVs, is R B = 1 − R M .

Authorized licensed use limited to: XIDIAN UNIVERSITY. Downloaded on December 09,2024 at 13:00:07 UTC from IEEE Xplore. Restrictions apply.
1890 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 19, 2024

Fig. 5. An identification sample of the proposed approach against distributed

spoofing attack in a UAV swarm with N = 30 UAVs and M = 4 malicious
UAVs, where d = 0.3. The dashed lines exhibit the misreported distance. Fig. 7. An identification sample of the proposed approach against mixed
spoofing attack, in a UAV swarm with N = 30 UAVs and M = 6 malicious
UAVs, where d = 0.3. The dashed lines exhibit the misreported distance.

Fig. 6. An identification sample of the proposed approach against collusion

spoofing attack, in a UAV swarm with N = 30 UAVs and M = 4 malicious Fig. 8. The variations of malicious ratio.
UAVs, where d = 0.3. The dashed lines exhibit the misreported distance.

C. Resistance to Distributed Spoofing Attacks

B. Visualization of Proposed Malicious UAV Detection
Fig. 8 plots the malicious ratio and average available
Fig. 5 shows the identifications of malicious UAVs in a degree across different numbers of malicious UAVs, where
UAV swarm with N = 30 UAVs under distributed spoofing the average of 100 independently randomly generated swarms
attacks. There are four malicious UAVs launching distributed with consistent parameters with Fig. 5 is plotted. It is noticed
spoofing attacks. Both the proposed CDI and E-CDI algo- that the malicious ratio is nearly linear to the number of
rithms are simulated. Fig. 6 shows the identifications of malicious UAVs. When there are seven malicious UAVs,
malicious UAVs in the 30-UAV swarm under collusion attacks, the potential malicious ratio can reach zero, highlighting the
where four malicious UAVs conspire collusion attacks towards presence of multiple malicious UAVs can severely disrupt or
a benign UAV. In Fig. 6, we only simulate the proposed even dismantle the normal operation of a swarm.
E-CDI algorithm since the CDI algorithm is inapplicable to In Fig. 9(a), we observe the Precision of the proposed
collusion attacks, as discussed in Section IV-C. Fig. 7 shows algorithms compared to the benchmark methods. It is evident
the identifications of malicious UAVs in the 30-UAV swarm that both the CDI and E-CDI algorithms outperform the others
under mixed attacks, where three malicious UAVs conspire significantly. The difference in Precision between the two
collusion attacks towards a benign UAV, while three other algorithms can be attributed to the strategy they employ for
malicious UAVs launch distributed attacks. We only run the handling neighboring index sets. The E-CDI algorithm, which
proposed E-CDI algorithm in Fig. 7 for the same reason as selects multiple indices simultaneously, may introduce some
considered for Fig. 6. From Figs. 5, 6, and 7, we can see that redundancies, leading to a slightly inferior Precision compared
the proposed CDI and E-CDI algorithms can effectively detect to the CDI algorithm, which selects one index at a time.
and identify malicious UAVs in applicable scenarios. Additionally, the decreasing trend in performance of both

Authorized licensed use limited to: XIDIAN UNIVERSITY. Downloaded on December 09,2024 at 13:00:07 UTC from IEEE Xplore. Restrictions apply.
BI et al.: DETECTION AND MITIGATION OF POSITION SPOOFING ATTACKS 1891

Fig. 9. (a) The performance on precision of the proposed and baseline approaches. (b) The performance on recall of the proposed and baseline approaches.
(c) The performance on F1 of the proposed and baseline approaches. (d) The performance on F1 of the proposed and baseline approaches under different
scales of the swarm network. (e) The performance on F1 of the proposed and baseline approaches under different levels of distance measurement noise.
(f) The performance on F1 of the proposed and baseline approaches under different levels of measurement distance.

proposed approaches can be linked to the degradation of the UAVs, where the proposed CDI and E-CDI algorithms con-
network structure, as indicated by the increasing malicious sistently outperform all other algorithms. It is noticed that the
ratio in Fig. 8. E-CDI algorithm is initially better than the CDI algorithm.
On the other hand, there is a noticeable upward trend in the However, as the number of malicious UAVs increases, the CDI
detection of malicious UAVs under the NLOS-based approach, algorithm gradually surpasses the E-CDI algorithm when the
especially those with significant distance errors akin to NLOS number of malicious UAVs exceeds five. The reason is that
conditions. The NLOS-based approach, which involves select- when there are only a small number of malicious UAVs in
ing UAVs with the largest distance errors, simultaneously a swarm, the E-CDI algorithm is more likely to misclassify
amplifies the likelihood of encountering malicious UAVs. The benign UAVs as malicious. As the number of malicious UAVs
ascending trend in the Random approach can be explained increases and the network structure deteriorates, the E-CDI
by the situation where the rate of capturing malicious UAVs algorithm is increasingly advantageous. With a higher number
surpasses the expansion rate of the malicious set. of malicious UAVs, the E-CDI algorithm exhibits a greater
In Fig. 9(b), we examine the Recall of the proposed methods likelihood of correctly detecting malicious UAVs.
compared to the benchmarks across varying numbers of mali- We also notice in Fig. 9(c) that the NLOS-based approach
cious UAVs. Both the CDI and E-CDI algorithms outperform outperforms the Random approach due to its selection from
all benchmarks significantly. Moreover, E-CDI offers higher a relatively smaller malicious set with a higher probability,
Recall than CDI. The reason is that E-CDI sacrifices some achieved through sorting based on absolute distance error.
precision, potentially introducing redundancies, to ensure the In contrast, the Random approach selects from a larger set
capture of more malicious UAVs. Given the critical security encompassing all possible malicious UAVs. The increasing
nature of the problem studied, it is imperative to emphasize trend observed in both benchmark algorithms is attributed to
high Recall to identify as many malicious UAVs as possible the growing number of malicious UAVs, resulting in a higher
to safeguard the swarm. As also noticed, the Recall of both probability of detection by both benchmarks.
proposed algorithms declines due to the deteriorating network Fig. 9(d) presents the F1 performance of our proposed
structure, as discussed in Fig. 9(a). For the same reason, the algorithms compared to the benchmark algorithms across
Recall of the NLOS-based approach also declines, as discussed these varying network scales. Both the proposed CDI and E-
in Fig. 9(a). As for the Random-based approach, it is intriguing CDI algorithms consistently outperform the other algorithms.
to note that the trends in Precision in Fig. 9(a) and Recall in An intriguing observation is that the CDI algorithm surpasses
Fig. 9(b) bear resemblance. The conclusion drawn is that the the E-CDI algorithm across all network scales. This is primar-
expansion rate of captured malicious UAVs exceeds that of the ily attributed to the fixed number of malicious UAVs: As the
malicious set enlargement, leading to this consistent trend. network scales up, a larger number of benign UAVs are likely
Fig. 9(c) illustrates the F1 of the proposed methods along- to be present, offering supporting evidence. However, this
side the benchmarks across varying numbers of malicious also introduces a risk for the E-CDI algorithm, which, while

Authorized licensed use limited to: XIDIAN UNIVERSITY. Downloaded on December 09,2024 at 13:00:07 UTC from IEEE Xplore. Restrictions apply.
1892 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 19, 2024

aiming for higher Recall, may inadvertently incorporate benign

UAVs. By contrast, the CDI algorithm, which emphasizes
higher Precision by meticulously identifying one UAV at a
time, mitigates this risk. On the other hand, the declining trend
observed in both benchmarks can be attributed to the larger
network scale, leading to a lower probability of detection
for both benchmarks. In particular, the increased number of
malicious UAVs in a larger-scale network makes it hard for
the benchmarks, which employ methods like selecting the
largest absolute distance error or sampling from the potential
malicious set, to identify the malicious UAVs effectively.
Fig. 9(e) presents the F1 performance of the proposed algo-
rithms in comparison to the benchmarks across different levels
of distance measurement inaccuracies. It is evident that all the Fig. 10. The number of iterations required by the CDI and E-CDI algorithms.
algorithms exhibit a consistent declining trend, resulting from
the adverse influence of inaccurate distance measurements
on the effectiveness of constraints in (12). Nevertheless, the advantageous. But, as the number of malicious UAVs expands,
proposed algorithms maintain efficiency and robustness even the escalating number of malicious UAVs poses increasing
in the face of elevated levels of measurement noise. This feasibility challenges for the CDI algorithm. On the other
resilience stems from the fact that the algorithms leverage hand, by prioritizing high Recall, the E-CDI algorithm aims
the entire swarm to aggregate evidence, effectively mitigating to identify as many malicious UAVs as possible.
the impact of inaccurate distance measurements. On the other
hand, the benchmarks face challenges with an increase in
noise, as these inaccuracies can lead to greater disparities D. Resistance to Collusion Spoofing Attacks
between the distances computed based on reported positions In line with Fig. 6, we delve into the evaluation of the F1
and the reported distance measurements. This, in turn, expands metric for the proposed E-CDI algorithm in comparison to the
the candidate set of malicious UAVs, reducing the likelihood benchmarks. The CDI algorithm is not assessed, as explained
of detection by the benchmarks. in Section IV. To maintain consistency, we generate 100
Fig. 9(f) provides an overview of the F1 performance of UAV swarms randomly and independently with consistent
the proposed algorithms and the benchmarks across different parameters with those in Fig. 6, except for the number of
measurement distances. In the case of the proposed algorithms, malicious UAVs. Each data point represents the average of
both initially exhibit an ascending trend, followed by a descent the 100 swarms.
after reaching a threshold of 0.35. This is because the increase Fig. 11(a) compares the F1 score between the E-CDI
in measurement distance prompts more UAVs to become algorithm and the benchmarks under a collusion spoofing
neighbors, thereby generating additional evidence for decision- attack. Unlike the declining trend observed in Fig. 9(c),
making at the beginning. However, the larger measurement the proposed E-CDI algorithm exhibits an upward trend as
distance relaxes the constraints applied to directly counter- malicious UAVs increase. On the one hand, collusion spoof-
spoofing, permitting more malicious UAVs to evade detection ing attacks have a distinct structure compared to distributed
and the F1 declines. In contrast, when considering the bench- spoofing attacks in the sense that attackers are more likely
mark algorithms, the expansion of measurement distance to be densely concentrated around a targeted UAV. On the
implies that more neighbors are involved in the evidence- other hand, the E-CDI algorithm assesses potentially malicious
gathering process. This can result in a higher number of UAVs individually if their neighborhood is detected to be
UAVs reporting pairwise distance measurements, consequently infected by malicious UAVs, which is particularly effective
leading to greater disparities between the distances computed in the densely populated neighborhood of a collusion attack.
based on reported positions and the reported distance measure- Consequently, the E-CDI can exploit the specific characteris-
ments. The resulting larger candidate set of malicious UAVs tics of collusion spoofing attacks.
reduces the probability of detection by the benchmarks. The ascending trend is also observed under the benchmark
We further compare the proposed CDI and E-CDI algo- algorithms, resulting from the presence of densely concen-
rithms in the number of iterations, as shown in Fig. 10. All trated UAVs, which tend to induce fewer disparities between
parameters are kept consistent with those in Fig. 5, except for the distance computed based on reported positions and the
the number of malicious UAVs. When the number of malicious reported distance measurements. This, in turn, results in
UAVs is relatively small, the CDI algorithm outperforms the smaller candidate sets of malicious UAVs, which are more
E-CDI algorithm. As the number of malicious UAVs increases, likely to be detected by the benchmarks.
a shift unfolds: the E-CDI algorithm gradually overtakes We proceed to assess the impact of two critical parame-
the CDI algorithm, and the performance gap between them ters: distance measurement error and the range for distance
widens with the growing number of malicious UAVs. This measurement on the performance of the proposed E-CDI
is because, when dealing with a small number of malicious algorithm. With the increasing distance measurement noises,
UAVs, the CDI algorithm’s scrutiny of each individual UAV is a descending trend of the F1 score can be observed in

Authorized licensed use limited to: XIDIAN UNIVERSITY. Downloaded on December 09,2024 at 13:00:07 UTC from IEEE Xplore. Restrictions apply.
BI et al.: DETECTION AND MITIGATION OF POSITION SPOOFING ATTACKS 1893

Fig. 11. The performance on F1 of the proposed and baseline approaches in collusion spoofing attack scenario. (a) Under different scales of malicious UAVs.
(b) Under different levels of distance measurement noise. (c) Under different levels of measurement distance.

Fig. 12. The performance evaluation on F1 of the proposed and baseline approaches in mixed spoofing attack scenario. (a) Under different scales of malicious
UAVs. (b) Under different levels of distance measurement noise. (c) Under different levels of measurement distance.

Fig. 11(b), as done in Fig. 9(e). On the other hand, a descend- Comparing the trends seen in Fig. 12(a) to those in Figs. 9(c)
ing trend is also noticed with the increasing permissible and 11(a), it is evident that the performance of the proposed
distance in Fig. 11(c), which is consistent with the observation E-CDI algorithm follows a similar ascending trend as observed
made in Fig. 9(f). Given that these two parameters exhibit a in Fig. 11(a), which is different from the trend in Fig. 9(c).
similar influence in both distributed and collusion scenarios, As malicious UAVs increase, the network structure undergoes
we can refer to the discussions about distributed spoofing in a notable transformation. Initially, both distributed attacks and
Section V-C for the sake of brevity. collusion attacks contribute evenly. However, with a greater
number of attackers, more malicious UAVs initially involved
E. Resistance to Mixed Spoofing Attack in distributed attacks inadvertently become participants in
As considered in Fig. 7, we proceed to assess the F1 collusion attacks. This shift results in the gradual dominance
metric of the proposed E-CDI algorithm, comparing it to of collusion attacks. Consequently, the performance trend
the benchmarks. A total of 100 UAV swarms are generated observed in the mixed spoofing attack aligns with the pattern
randomly and independently with consistent parameters with shown in Fig. 11(a), although there can be a slight performance
those considered in Fig. 7, except for the number of mali- degradation for the same number of malicious UAVs.
cious UAVs. Each data point represents the average of the In the case of the benchmarks, the ascending trend of
100 swarms. their F1 scores can be attributed to the presence of densely
In Fig. 12(a), we analyze the F1 performance of the concentrated UAVs, which tends to reduce the disparities
proposed E-CDI algorithm compared to the benchmarks in between the distance computed based on reported positions
the context of a mixed spoofing attack. In order to con- and the reported distance measurements, as discussed earlier.
duct a fair evaluation across various numbers of malicious Last but not least, we assess the impact of distance mea-
UAVs, we evenly distribute the malicious UAVs into two surement error and the permissible distance for distance
distinct groups. For instance, when dealing with six malicious measurement on the performance of the proposed E-CDI
UAVs, we assign three of them to execute distributed attacks. algorithm. It is noticed that Fig. 12(b) yields a declining
The remaining three are directed toward launching collusion trend like the one observed in Fig. 9(e), while Fig. 12(c)
attacks against a benign UAV. It is noted that the mixed displays a decreasing pattern like the one shown in Fig. 9(f).
spoofing attack cannot be simply regarded as the superposition Given that these two parameters have consistent effects under
of distributed and collusion attacks. The amalgamation of these the distributed and collusion attacks, the reason underlying
different attack types within the mixed scenario introduces the observations in Figs. 12(b) and 12(c) can be established,
substantial variations in the network structure. as discussed in Section V-C. It is worth highlighting that there

Authorized licensed use limited to: XIDIAN UNIVERSITY. Downloaded on December 09,2024 at 13:00:07 UTC from IEEE Xplore. Restrictions apply.
1894 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 19, 2024

exists a marginal performance decline for the same number [16] B. Pardhasaradhi and L. R. Cenkeramaddi, “GPS spoofing detection and
of malicious UAVs when compared to the distributed attacks. mitigation for drones using distributed radar tracking and fusion,” IEEE
Sensors J., vol. 22, no. 11, pp. 11122–11134, Jun. 2022.
This is due to the more intricate degradation of the network [17] A. Khan, N. Ivaki, and H. Madeira, “Are UAVs’ flight controller software
structure exacerbated by the interplay of mixed attacks. reliable?” in Proc. IEEE PRDC, 2022, pp. 194–204.
[18] K. Liu, Y. Zhao, G. Wang, and R. Xue, “Quadrotor visual navigation
VI. C ONCLUSION under GPS spoofing attack,” in Proc. 6th Int. Conf. Wireless Commun.,
Signal Process. Netw. (WiSPNET), Mar. 2021, pp. 270–277.
In this paper, we judiciously formulated a complicated
[19] Y. Dang, C. Benzaïd, T. Taleb, B. Yang, and Y. Shen, “Transfer learning
malicious UAV detection problem based on the reported based GPS spoofing detection for cellular-connected UAVs,” in Proc. Int.
positions and pairwise distance measurements of the UAVs as Wireless Commun. Mobile Comput. (IWCMC), May 2022, pp. 629–634.
a localization feasibility problem. Then, we relied on an SDR [20] W. Chen, Y. Dong, and Z. Duan, “Accurately redirecting a malicious
drone,” in Proc. IEEE 19th Annu. Consum. Commun. Netw. Conf.
approach to recast the formulated non-convex problem into a (CCNC), Jan. 2022, pp. 827–834.
convex SDP, which is the key to assessing the feasibility of [21] Z. Wang, X. Zhang, Z. Zhou, M. Lu, and H. Li, “GNSS spoofer
the reported positions and distance measurements. Moreover, localization for vehicles based on Doppler and clock drift double
difference,” IEEE Trans. Veh. Technol., vol. 72, no. 4, pp. 4466–4481,
we proposed two tailored iterative algorithms based on the Apr. 2023.
proposed SDP approach to detect and identify malicious UAVs [22] G. Michieletto, F. Formaggio, A. Cenedese, and S. Tomasin, “Robust
in UAV swarms. Extensive simulations demonstrated that the localization for secure navigation of UAV formations under GNSS
spoofing attack,” IEEE Trans. Autom. Sci. Eng., vol. 20, no. 4,
proposed algorithms can achieve superior performance to the pp. 2383–2396, Oct. 2023.
existing benchmarks and exhibit robustness across various [23] M. Babaghayou, N. Labraoui, A. A. A. Ari, N. Lagraa, M. A. Ferrag, and
UAV swarms. L. Maglaras, “SAMA: Security-aware monitoring approach for location
abusing and UAV GPS-spoofing attacks on Internet of Vehicles,” in Proc.
R EFERENCES LNICST, vol. 427, 2022, pp. 343–360.
[24] S. Bi, J. Cui, W. Ni, Y. Jiang, S. Yu, and X. Wang, “Three-dimensional
[1] S. Hu, W. Ni, X. Wang, A. Jamalipour, and D. Ta, “Joint optimization of
cooperative positioning for Internet of Things provenance,” IEEE Inter-
trajectory, propulsion, and thrust powers for covert UAV-on-UAV video
net Things J., vol. 9, no. 20, pp. 19945–19958, Oct. 2022.
tracking and surveillance,” IEEE Trans. Inf. Forensics Security, vol. 16,
pp. 1959–1972, 2021. [25] Y. Wang, X. Wen, Y. Cao, C. Xu, and F. Gao, “Bearing-based relative
localization for robotic swarm with partially mutual observations,” IEEE
[2] B. Liu, W. Ni, R. P. Liu, Y. J. Guo, and H. Zhu, “Decentralized, privacy-
Robot. Autom. Lett., vol. 8, no. 4, pp. 2142–2149, Apr. 2023.
preserving routing of cellular-connected unmanned aerial vehicles for
joint goods delivery and sensing,” IEEE Trans. Intell. Transp. Syst., [26] J. Yang, T. Zhang, X. Wu, T. Liang, and Q. Zhang, “Efficient scheduling
vol. 24, no. 9, pp. 9627–9641, 2023. in space–air–ground-integrated localization networks,” IEEE Internet
Things J., vol. 9, no. 18, pp. 17689–17704, Sep. 2022.
[3] X. Yuan, S. Hu, W. Ni, X. Wang, and A. Jamalipour, “Deep reinforce-
ment learning-driven reconfigurable intelligent surface-assisted radio [27] B. Jiang, B. D. O. Anderson, and H. Hmam, “3-D relative localization
surveillance with a fixed-wing UAV,” IEEE Trans. Inf. Forensics Secu- of mobile systems using distance-only measurements via semidefinite
rity, vol. 18, pp. 4546–4560, 2023. optimization,” IEEE Trans. Aerosp. Electron. Syst., vol. 56, no. 3,
pp. 1903–1916, Jun. 2020.
[4] A. V. Savkin, W. Ni, and M. Eskandari, “Effective UAV navigation for
cellular-assisted radio sensing, imaging, and tracking,” IEEE Trans. Veh. [28] Z. Wu, W. Quan, and T. Zhang, “Resource allocation in UAV-aided
Technol., vol. 72, no. 10, pp. 13729–13733, Oct. 2023. vehicle localization frameworks,” in Proc. IEEE/CIC Int. Conf. Commun.
Workshops China, Aug. 2019, pp. 98–103.
[5] K. Li, R. C. Voicu, S. S. Kanhere, W. Ni, and E. Tovar, “Energy efficient
legitimate wireless surveillance of UAV communications,” IEEE Trans. [29] P. Biswas, T.-C. Lian, T.-C. Wang, and Y. Ye, “Semidefinite pro-
Veh. Technol., vol. 68, no. 3, pp. 2283–2293, Mar. 2019. gramming based algorithms for sensor network localization,” ACM
Trans. Sensor Netw., vol. 2, no. 2, pp. 188–220, May 2006, doi:
[6] S. Hu, X. Yuan, W. Ni, and X. Wang, “Trajectory planning of cellular-
10.1145/1149283.1149286.
connected UAV for communication-assisted radar sensing,” IEEE Trans.
[30] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.:
Commun., vol. 70, no. 9, pp. 6385–6396, Sep. 2022.
Cambridge Univ. Press, 2004.
[7] X. Yuan, Z. Feng, W. Ni, R. P. Liu, J. A. Zhang, and W. Xu,
[31] M. Grant and S. Boyd. (2008). CVX: MATLAB Software for Disciplined
“Secrecy performance of terrestrial radio links under collaborative
Convex Programming. [Online]. Available: http://cvxr.com/cvx
aerial eavesdropping,” IEEE Trans. Inf. Forensics Security, vol. 15,
pp. 604–619, 2020. [32] I. Um, S. Park, S. Oh, and H. Kim, “Analyzing location accu-
racy of unmanned vehicle according to RTCM message frequency
[8] B. Song et al., “Reliability analysis of large-scale adaptive weighted net-
of RTK-GPS,” in Proc. 25th Asia–Pacific Conf. Commun. (APCC),
works,” IEEE Trans. Inf. Forensics Security, vol. 15, pp. 651–665, 2020.
Nov. 2019, pp. 326–330.
[9] X. Yuan, Z. Feng, W. Ni, Z. Wei, R. P. Liu, and C. Xu, “Connectivity
[33] W. Zhao, R. He, B. Ai, Z. Zhong, and H. Zhang, “Vehicle localization
of UAV swarms in 3D spherical spaces under (un)intentional ground
based on hypothesis test in NLOS scenarios,” IEEE Trans. Veh. Technol.,
interference,” IEEE Trans. Veh. Technol., vol. 69, no. 8, pp. 8792–8804,
vol. 71, no. 2, pp. 2198–2203, Feb. 2022.
Aug. 2020.
[34] S. Bi, W. Ni, Y. Jiang, and X. Wang, “Novel recommendation-based
[10] S. Bi et al., “A comprehensive survey on applications of AI technologies
approach for multidisciplinary development of future universities,” Sus-
to failure analysis of industrial systems,” Eng. Failure Anal., vol. 148,
tainability, vol. 14, no. 10, p. 5881, May 2022.
Jun. 2023, Art. no. 107172.
[11] Y. E. Yao, P. Dash, and K. Pattabiraman, “May the swarm be with you:
Sensor spoofing attacks against drone swarms,” in Proc. ACM SIGSAC
CCS, 2022, pp. 3511–3513.
[12] H. Choi et al., “Detecting attacks against robotic vehicles: A control
invariant approach,” in Proc. ACM SIGSAC Conf. Comput. Commun.
Secur., Oct. 2018, pp. 801–816. Siguo Bi received the Ph.D. degree in electronic
[13] I. Buske, A. Walther, D. Fitz, J. Acosta, A. Konovaltsev, and L. Kurz, science and technology from Fudan University,
“Smart GPS spoofing to countermeasure autonomously approaching Shanghai, China, in 2020. He is currently an Engi-
agile micro UAVs,” Proc. SPIE, vol. 12273, pp. 62–67, Nov. 2022. neer with Fudan University. His research interests
[14] Y. Liu, Y. Wang, J. Wang, and Y. Shen, “Distributed 3D relative include wireless sensor networks, convex optimiza-
localization of UAVs,” IEEE Trans. Veh. Technol., vol. 69, no. 10, tion, network security in UAV swarm, and machine
pp. 11756–11770, Oct. 2020. learning applied in communications.
[15] P. Jiang, H. Wu, and C. Xin, “DeepPOSE: Detecting GPS spoofing attack
via deep recurrent neural network,” Digit. Commun. Netw., vol. 8, no. 5,
pp. 791–803, Oct. 2022.

Authorized licensed use limited to: XIDIAN UNIVERSITY. Downloaded on December 09,2024 at 13:00:07 UTC from IEEE Xplore. Restrictions apply.
BI et al.: DETECTION AND MITIGATION OF POSITION SPOOFING ATTACKS 1895

Kai Li (Senior Member, IEEE) received the B.E. since 2018 and an Editor for IEEE T RANSACTIONS ON V EHICULAR T ECH -
degree from Shandong University, Weihai, China, NOLOGY . He served as the Secretary, then the Vice-Chair and Chair of the
in 2009, the M.S. degree from The Hong Kong IEEE Vehicular Technology Society (VTS) New South Wales (NSW) Chapter
University of Science and Technology, Hong Kong, from 2015 to 2022, the Workshop Chair for ISCIT 2023, the Track Chair
in 2010, and the Ph.D. degree in computer science for VTC-Spring 2017, the Track Co-Chair for IEEE VTC-Spring 2016, the
from The University of New South Wales, Sydney, Publication Chair for BodyNet 2015, and the Student Travel Grant Chair for
NSW, Australia, in 2014. Currently, he is a Visiting WPMC 2014.
Research Scientist with the Division of Electrical
Engineering, Department of Engineering, University
of Cambridge, U.K. Additionally, he is also a Senior
Research Scientist with the CISTER Research Cen-
tre, Porto, Portugal. He is also a CMU-Portugal Research Fellow, jointly
supported by Carnegie Mellon University, Pittsburgh, PA, USA, and the
Foundation for Science and Technology (FCT), Lisbon, Portugal. In 2022,
he was a Visiting Research Scholar with the CyLab Security and Privacy
Institute, CMU. Prior to this engagement, he was a Post-Doctoral Research Cong Wang received the B.E. and M.S. degrees
Fellow with the SUTD-MIT International Design Centre, Singapore University from Fudan University, China. Her research inter-
of Technology and Design, Singapore, from 2014 to 2016. He has also held ests include machine learning, bioinformatics, and
positions as a Visiting Research Assistant with the ICT Centre, CSIRO, statistics.
Brisbane, QLD, Australia, from 2012 to 2013, and a Research Assistant with
the Mobile Technologies Centre, The Chinese University of Hong Kong,
Hong Kong, from 2010 to 2011. He has been an Associate Editor for
the Nature Computer Science journal (Springer) since 2023, the Computer
Communications journal (Elsevier) and Ad Hoc Networks journal (Elsevier)
since 2021, and IEEE ACCESS journal since 2018.

Shuyan Hu (Member, IEEE) received the B.Eng.

degree in electrical engineering from Tongji Uni-
versity, China, in 2014, and the Ph.D. degree
in electronic science and technology from Fudan
University, China, in 2019. She is currently a
Post-Doctoral Research Fellow with the School of
Information Science and Technology, Fudan Univer-
sity. She was selected by the Shanghai Post-Doctoral Xin Wang (Fellow, IEEE) received the B.Sc. and
Excellence Program in 2019. Her research interests M.Sc. degrees in electrical engineering from Fudan
include machine learning and convex optimizations University, Shanghai, China, in 1997 and 2000,
and their applications to unmanned aerial vehicle respectively, and the Ph.D. degree in electrical engi-
(UAV) networks and intelligent systems. neering from Auburn University, Auburn, AL, USA,
in 2004.
Wei Ni (Fellow, IEEE) received the B.E. and From September 2004 to August 2006, he was a
Ph.D. degrees in electronic engineering from Fudan Post-Doctoral Research Associate with the Depart-
University, Shanghai, China, in 2000 and 2005, ment of Electrical and Computer Engineering,
respectively. University of Minnesota, Minneapolis. In August
Currently, he is a Principal Research Scientist 2006, he joined the Department of Electrical Engi-
with CSIRO, Sydney, Australia, a Conjoint Pro- neering, Florida Atlantic University, Boca Raton, FL, USA, as an Assistant
fessor with the University of New South Wales, Professor, then was promoted to a tenured Associate Professor in 2010.
an Adjunct Professor with the University of Tech- He is currently a Distinguished Professor and the Chair of the Department
nology Sydney, and an Honorary Professor with of Communication Science and Engineering, Fudan University. His research
Macquarie University. He was a Post-Doctoral interests include stochastic network optimization, energy-efficient communica-
Research Fellow with Shanghai Jiao Tong Univer- tions, cross-layer design, and signal processing for communications. He served
sity from 2005 to 2008, the Deputy Project Manager with Bell Labs, as a Senior Area Editor for IEEE T RANSACTIONS ON S IGNAL P ROCESSING,
Alcatel/Alcatel-Lucent from 2005 to 2008, and a Senior Researcher with an Associate Editor for IEEE T RANSACTIONS ON S IGNAL P ROCESSING,
Devices Research and Development, Nokia, from 2008 to 2009. He has an Editor for IEEE T RANSACTIONS ON W IRELESS C OMMUNICATIONS,
authored eight book chapters, over 300 journal articles, 100 conference an Editor for IEEE T RANSACTIONS ON V EHICULAR T ECHNOLOGY, and an
papers, 26 patents, and ten standard proposals accepted by IEEE. His research Associate Editor for IEEE S IGNAL P ROCESSING L ETTERS. He is a member
interests include machine learning, online learning, stochastic optimization, of the Signal Processing for Communications and Networking Technical
and their applications to system efficiency and integrity. He has been serving Committee of IEEE Signal Processing Society and a Distinguished Speaker
as an Editor for IEEE T RANSACTIONS ON W IRELESS C OMMUNICATIONS of the IEEE Vehicular Technology Society.

Authorized licensed use limited to: XIDIAN UNIVERSITY. Downloaded on December 09,2024 at 13:00:07 UTC from IEEE Xplore. Restrictions apply.