Computers & Industrial Engineering 196 (2024) 110454

Contents lists available at ScienceDirect

Computers & Industrial Engineering

journal homepage: www.elsevier.com/locate/caie

A swarm anomaly detection model for IoT UAVs based on a multi-modal

denoising autoencoder and federated learning
Yu Lu a , Tao Yang a, * , Chong Zhao a , Wen Chen b , Rong Zeng c
a
School of Computer Science, China West Normal University, Nanchong, China
b
the School of Cyber Science and Engineering, Sichuan University, Chengdu, China
c
School of Electronic Information Engineering, China West Normal University, Nanchong, China

A R T I C L E I N F O A B S T R A C T

Keywords: The widespread application of unmanned aerial vehicle (UAV) swarms has posed unique challenges for anomaly
UAV swarm detection. Multi-modal noise from multi-source heterogeneous sensors during UAV swarm communication af-
Intrusion detection fects data quality, and limited data sharing between different UAV organisations restricts training a unified
Federated learning
anomaly detection model. To address these problems, this study proposes a UAV swarm anomaly detection
Denoising autoencoder
model based on a multi-modal denoising autoencoder and federated learning (L-MDAE). First, L-MDAE simulates
noise by adding perturbations to the original data during UAV swarm communication. Second, according to the
characteristics of UAV data noise, this study designs a new MSE loss function (normalised mean square error,
NMSE) based on the normalised correlation coefficient. Furthermore, heterogeneous neural networks with NMSE
are constructed to enhance the multi-modal noise–removal capability of the model. Finally, this study considers
the UAV control node as the client and the ground control station as the server. Using a federated learning
mechanism, L-MDAE is trained on a client dataset, and its parameters are integrated and distributed on the
server. In this way, each UAV can effectively detect abnormal data using L-MDAE. Experimental results on five
datasets, including ALFA, TLM and ITS, demonstrate that L-MDAE outperforms baseline and related models.
When using ALFA, L-MDAE achieved an accuracy of 0.9919 and a swarm anomaly detection accuracy of 0.9901,
approximately 2% higher than that of the baseline model.

1. Introduction Fig. 1 shows several typical network topologies for UAV swarms.
Fig. 1(a) illustrates a UAV network architecture involving individual
With the development of control, navigation and communication UAVs and a GCS. UAVs can directly communicate with the GCS. Fig. 1(b)
technology, unmanned aerial vehicles (UAVs) and their swarms have shows the multi-UAV star topology. In this network structure, individual
been successfully applied in agriculture (Albiero et al., 2022), intelligent UAVs do not have direct data exchange. Each UAV can directly
cities (Hayawi et al., 2023), environmental monitoring (Motlagh et al., communicate with the GCS, while communication between UAVs re-
2023) and rescue operations (Horyna et al., 2023). As UAVs and their quires assistance of the GCS. Fig. 1(c) shows a multi-UAV self-organising
swarms exhibit increasingly widespread applications, the demand for swarm network, typically consisting of two or more swarm groups. Each
enhanced reliability and stability increases. The communication archi- group of UAVs forms a network, with one UAV in each group acting as a
tecture of UAV swarms can be divided into three parts: ground control central node responsible for data exchange with the GCS.
stations (GCSs), individual UAVs and communication links. Based on the As shown in Fig. 1, in a multi-UAV swarm working environment,
number of UAVs in a swarm, the UAV network can be divided into a owing to the communication links among UAVs and between the UAV
single UAV network and a multi-UAV network. Depending on the central node and the GCS, communication noise is generated between
different communication methods between UAVs and GCSs, the multi- these links in this multi-swarm environment, further causing signal
UAV network can be further divided into a multi-UAV star network to- noise to the sensors of individual UAVs. In addition, UAVs are aircraft
pology and a multi-UAV self-organising grouping network topology. with multiple sensors, and the communication interferences mentioned

* Corresponding author.
E-mail addresses: luyu@stu.cwnu.edu.cn (Y. Lu), yangt@cwnu.edu.cn (T. Yang), 79335604@qq.com (C. Zhao), wenchen@scu.edu.cn (W. Chen), 410012zr@
cwnu.edu.cn (R. Zeng).

https://doi.org/10.1016/j.cie.2024.110454
Received 18 April 2024; Received in revised form 19 July 2024; Accepted 3 August 2024
Available online 13 August 2024
0360-8352/© 2024 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
Y. Lu et al. Computers & Industrial Engineering 196 (2024) 110454

Fig. 1. UAV swarm network topology.

above can result in different sensors of UAVs experiencing multi-modal of UAV swarms.
noise. For a detailed problem analysis, refer to Section 4.1. To address the above-mentioned limitations, this study proposes a
The UAV swarm will likely experience malfunctions owing to com- model for anomaly detection in IoT UAV swarms based on a multi-modal
plex natural environments, flight tasks, signal noise or human-made denoising autoencoder and federated learning (L-MDAE). First, by
cyberattacks. For example, in Hong Kong, GPS interference caused a adding noise to the UAV on-board abnormal data, L-MDAE simulates the
swarm of UAVs to crash during a demonstration over Victoria Harbour, data noise generated by swarming flights. Second, a loss function called
resulting in a loss of 10,000 Hong Kong dollars (South China Morning the normalised mean square error (NMSE) was designed using the MSE
Post, 2018). In addition, critical components of UAVs may malfunction and normalised correlation coefficient (NCC). NMSE allows L-MDAE to
in complex natural environments, such as complex weather conditions, adjust its weight parameters based on the NCC and MSE, effectively
leading to the crash of UAVs and thus posing a threat to human life and fitting the clean data. Moreover, a multi-modal denoising autoencoder
property safety (Air Accidents Investigation Branch reports, 2020; Seo (DAE) was designed based on the characteristics of multi-source het-
et al., 2015). The University of Oxford in the UK (Brundage et al., 1802; erogeneous sensors of UAVs. The L-MDAE model uses a multi-modal
Replicating peregrine attack strategies could help down rogue UAVs), neural network to extract features from heterogeneous UAV sensor
the United Arab Emirates (Al Amir et al., 2018; Asadzadeh et al., 2022; data to enhance its denoising capability. Finally, the central node of the
Johnsen et al., 2020) and the University of New South Wales in UAV was abstracted as a client for federated learning, while the GCS was
Australia, among other high education institutions and research orga- abstracted as the server for federated learning. The L-MDAE model was
nisations, have already demonstrated the potential vulnerabilities in trained using a pre-training method through unsupervised training,
UAV systems and proven the likelihood of them being attacked in where parameters were frozen after training to construct the denoising
various scenarios through experiments (Krishna & Murphy, 2017). model. A fully connected classification layer was added afterwards to
Anomaly detection is one of the primary technical means to ensure obtain the anomaly detection model.
the stability of UAV swarm flight. Many scholars have conducted Data were collected through the UAV control node and uploaded to
extensive research in this field (this is discussed in Section 2). Many the GCS. The GCS trained and distributed the model to the UAV control
machine learning models, such as KNN, SVM and decision trees, are node. Through federated learning, the training and deployment of the
applied to anomaly detection for UAVs and their swarms. In addition, UAV swarm anomaly detection model could be efficiently completed
deep learning models such as multilayer perceptron (MLP), convolu- while ensuring privacy and efficient data transmission.
tional neural networks (CNNs) and long short-term memory neural The main contributions of this study are as follows:
networks (LSTMs) are applied to anomaly detection for UAVs and their
swarms. However, the communication noise generated by the UAV • This study proposes a loss function based on the NCC and MSE that
swarm during operation causes multi-modal noise interference from can better remove noise, giving the denoising model a more robust
multiple sources to the UAV sensors. This phenomenon affects the clean data–fitting capability.
denoising capability of data-driven swarm anomaly detection models, • This study introduces a multi-modal DAE that utilises heterogeneous
limiting their anomaly detection performance. Furthermore, the lack of neural networks to extract noise features from multi-source hetero-
sensor data sharing among heterogeneous UAV swarms from multiple geneous UAV sensor data, enhancing the noise-removal capability of
agencies hinders their ability to collaboratively train a high- the model.
performance anomaly detection model, further reducing the stability

2
Y. Lu et al. Computers & Industrial Engineering 196 (2024) 110454

Fig. 2. Denoising autoencoder.

• This study integrates the improved loss function of the L-MDAE 2024, Ahn et al. (Ahn & Chung, 2024) proposed a UAV swarm anomaly
model for anomaly detection in a UAV swarm. We use the federated detection model based on semi-supervised learning. Using normal UAV
learning mechanism to abstract the UAV control node as the client flight data, the model trained various one-dimensional convolution-
and the GCS as the server. The UAV control node collects anomaly based anomaly detection models, including variational autoencoders,
data and trains the L-MDAE model. The GCS then integrates and AnoGAN, GANomaly and Skip-GANomaly. The optimal hyper-
distributes model parameters, ultimately deploying the model on a parameters of the model were determined by Bayesian optimisation, and
single-point UAV. experimental results reveal that the model has good accuracy.

The remainder of this paper is organised as follows. Section 2 pre- 2.2. Machine learning–based anomaly detection model for UAVs
sents the previous work related to this study. Section 3 presents the
preparatory knowledge related to this study. Section 4 presents the In 2020, Alos et al. (Alos et al., 2020) used a PCA-based method to
proposed L-MDAE model. Section 5 presents the experiments, including detect anomalies in UAVs during flight. Using PCA, they identified the
comparison and ablation experiments. Section 6 discusses the general- contributions of different variables, such as UAV attitude data and
isation and practical application feasibility of the model. Finally, Section navigation data, during abnormal flights. Experiments showed that this
7 concludes the entire study. method, through unsupervised learning, does not require a large amount
of labelled training data and has high accuracy. In 2020, Bronz et al.
2. Related work (Bronz et al., 2020:) used SVM to detect possible faults during UAV
flights. Experimental results indicated that their model exhibits high
2.1. Anomaly detection models for UAV swarms accuracy; however, owing to limited training data samples, the model
suffered from overfitting issues, further limiting its generalisation ca-
In 2015, Wang et al. (Wang et al., 2015) proposed a method for pabilities. In 2021, Ayhan et al. (Altinors et al., 2021) detected anom-
detecting anomalous aircraft in a swarm of UAVs, which was based on alies in UAV engine sound signals. Initially, they designed common UAV
common model parameters shared by normal UAVs. This model first faults such as propeller, eccentric disc and bearing faults using real UAV
identifies the model parameters of each aircraft in the swarm based on models. Then, they obtained engine sound signals in these fault sce-
input and output data using a sparse optimisation technique. Second, the narios. Finally, they used decision trees, support vector machines and
model detects the flying status of individual UAVs by identifying KNN algorithms to perform UAV engine fault detection. This study uti-
changes in specific model parameters. Finally, the effectiveness of this lised sound signals for anomaly detection, providing new insights for
method is demonstrated through simulation experiments. In 2022, He future researchers.
et al. (He et al., 2022) proposed a UAV edge network intrusion detection
algorithm based on CGAN and LSTM. To mitigate challenges arising 2.3. Deep learning–based anomaly detection model for UAVs
from variations in local UAV datasets, limited sample sizes and uneven
dataset distributions, the model incorporates LSTM into the training In 2022, Bell et al. (Bell et al., 2203) proposed a UAV sensor anomaly
process of the CGAN. He et al. also incorporated a distributed federated detection model based on stacked LSTM-AEs. This model utilises an
learning architecture with blockchain and differential privacy methods autoencoder of stacked LSTMs to capture anomaly patterns of UAV
to ensure the security of model parameter uploading and local data sensors; furthermore, the model comprises a dynamic threshold algo-
privacy. Experimental results show that the model performs well. In rithm and a weighted loss function. Experimental results demonstrate
2023, Subbarayalu V et al. (Subbarayalu & Vensuslaus, 2023) presented that these dynamic thresholds and weighted loss functions can enhance
a UAV swarm anomaly detection model based on a time probabilistic the performance of the model in terms of accuracy and detection speed.
automaton. This model trains normal UAV behaviour to identify In 2023, Yang et al. (Yang et al., 2023) presented an unsupervised UAV
abnormal states of UAVs. The model determines the behaviour of the anomaly detection model based on spatiotemporal correlations. The
UAV by monitoring the behaviour of individual UAVs in a swarm. In model first pre-processes UAV flight data using the Savitzky–Golay

3
Y. Lu et al. Computers & Industrial Engineering 196 (2024) 110454

filtering data processing technique to mitigate the noise in the original ∑N ( )

i=1 Loss X
̂ i , Xʹ
historical flight data. Feature subset selection based on correlation is θDe , θEn = arg min i (5)
conducted to reduce reliance on expert knowledge. Finally, the extrac- θDe ,θEn n
ted features are used as inputs to the designed stacked LSTM autoen-
coder model to achieve unsupervised detection of UAV flight data 3.2. Normalised correlation coefficient
anomalies. In 2024, Yang et al. (Yang et al., 2024) introduced a
spatiotemporal-correlated multivariate regression UAV flight data The NCC is widely used in signal processing and computer vision.
anomaly detection model. The model initially performs parameter se- The NCC captures the overall similarity between signal waveforms
lection using correlation analysis. Subsequently, CNN-BiLSTM is used as before and after denoising, rather than focusing on specific waveform
a feature extraction model to strengthen the learning of key fault fea- details. The calculation method of the NCC is shown in Eq. (6).
tures in UAVs through an attention mechanism. Residual filtering is ∑N
introduced to smooth residuals, and anomaly detection is achieved by n=1 As(n) × Ad(n)
NCC = √(̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑N ) ( ∑N )̅ (6)
comparing smoothed residuals with statistical thresholds. n=1 As (n) × n=1 Ad (n)
2 2

The above authors have made great contributions to the field of

drone anomaly detection. Through the review and analysis of the above In Eq. (6), As(n) refers to the original pure signal and Ad(n) refers to
literature, it can be found that traditional UAV and swarm anomaly the signal filtered by the denoising algorithm,NCC ∈ [ − 1, 1]. When the
detection models have overlooked the data noise caused by communi- value is − 1, it represents that two signals are negatively correlated;
cation interference during swarm flight, resulting in insufficient noise when the value is 0, it represents that two signals are uncorrelated; when
resistance for the models. Furthermore, on-board data from UAVs usu- the value is 1, it represents that two vectors are positively correlated.
ally come from multiple sensors. Traditional DAEs typically rely on Therefore, the larger the value of NCC, the stronger the correlation be-
single-modal models for denoising, making it challenging to handle tween the two signals, and vice versa, suggesting that the similarity
noise data from multiple heterogeneous UAV sensor sources. Finally, the between the two signals is weaker.
lack of sensor data sharing from heterogeneous UAV swarm data from
multiple agencies has led to the inability to train a unified data-driven 4. Proposed L-MDAE model
anomaly detection model. The above problems lead to the insufficient
anomaly detection capabilities of UAV swarms, limiting their further 4.1. Problem analysis
application in complex environments. Thus, this study proposes an IoT
UAV swarm anomaly detection model based on multi-modal DAEs and Noise refers to the random fluctuations in sensor output signals, such
federated learning, which is detailed in Section 4. Table 14 presents the as slow variations, drift and sudden changes. These fluctuations can
list of acronyms in the study. arise from external interference, temperature variations or mechanical
defects. Sensor noise is generally divided into two categories: internal
3. Preliminaries noise and external noise. Internal noise is caused by random motion of
the sensor, including thermal noise, quantum noise and amplifier circuit
3.1. DAE noise; external noise is caused by factors such as communication elec-
tromagnetic interference, environmental changes and mechanical vi-
The DAE (Vincent et al., 2008) is an unsupervised learning algorithm brations. For clarity, the noise generated by UAV sensors is defined in
for noise reduction. DAE is an encoder–decoder architecture that in- this study as additive noise. As shown in Eq. (7), the observed signal of
troduces noise into input data and attempts to reconstruct clean data the sensor is x(t), the true value of the sensor is s(t), the noise due to
from the noisy data through the encoder and decoder, achieving noise communication interference is n(t) and Δ is the deviation of the noise
removal. Compared to traditional autoencoders, DAE has stronger signal. Suppose the real signal is set as a sine wave signal, with an
robustness and generalisation capabilities due to the introduction of amplitude of A, an angular frequency of ω and an initial phase of φ, then
noise in the input data. its noise signal follows a Gaussian noise with the location parameter μ
Fig. 2 illustrates the structure of the DAE. In Fig. 2, X refers to the and scale parameter σ.
original clean data and X⌢ refers to the data after adding noise to X. Xʹ ⎧
⎪ x(t) = s(t) + Δn(t)
⎪
refers to the data after DAE reconstruction, X, X⌢, Xʹ ∈ Rn , and n rep- ⎨
s(t) = Asin(ωt + φ) (7)
resents the dimension of the vector. The DAE encodes the data using a ⎪
⎪ ( )
⎩
decoder via Eq. (1). In Eq. (2), σ refers to the activation function.WDe and n(t) N μ, σ 2
bDe are the weight matrix and bias of the decoder, respectively, and h
Denoising aims to recover the original signal s(t) from the UAV
refers to the encoded vector.
observation signal x(t). Taking the wavelet transform as an example, the
h = fθ ( X)
̂ (1) wavelet-based signal denoising task can be modelled as the following
mathematical problem:
fθ ( X)
̂ = σ(WDe × X
̂ + bDe ) (2) ⎧
⎪
⎪
⎪ x(t) = s(t) + Δn(t)
The DAE uses Eq. (3) to decode h and obtain Xʹ. In Eq. (4), WEn , bEn are ⎪
⎪
⎪
⎪ {( ) }
⎪
the weight matrix and bias of the decoder, respectively. ⎪
⎪
⎪ Φ = {x|x(t)
J
}, Ψ = span ψ jk j=1 , ϕJk
⎪
⎨
Xʹ = gθʹ (h) (3) Γ = {β|β : Φ→Ψ} (8)
⎪
⎪
⎪
⎪
⎪ xmaxopt = βmaxopt (x)
gθʹ (h) = σ (WEn × h + bEn ) (4) ⎪
⎪
⎪
⎪ βmaxopt = argmin(‖β(x)− s‖ )
⎪
⎪ β∈Γ
⎪
The DAE uses the objective function of Eq. (5) to optimise its ⎩
parameter matrix. In Eq. (5), θDe and θEn refer to the parameter matrices
of encoder and decoder in DAE, respectively, and N refers to the number As shown in Eq. (8), the signal denoising task based on the wavelet
of samples. transform can be mathematically represented as finding the best
approximation function for the true signal in the function space con-
sisting of wavelet mother functions, according to the defined loss mea-
surement method to achieve noise removal. In Eq. (8), βmaxopt (x)

4
Y. Lu et al. Computers & Industrial Engineering 196 (2024) 110454

Fig. 3. Multi-modal noise in multi-sensor UAVs.

proposes a denoising model (L-MDAE) based on the NMSE loss function

Table 1
and a multi-modal DAE to effectively remove noise. The model structure
Feature classification in ALFA.
is shown in Fig. 4.
Categorisation Feature name Fig. 4 shows the L-MDAE model and its structure diagram for
Features of time-series classes in field.orientation.x anomaly detection in UAV swarms, including the multi-modal denoising
datasets field.orientation.y phase and the UAV swarm anomaly detection phase. Specifically, this
field.orientation.z model can be systematically divided into three parts.
field.orientation.w
field.linear_acceleration.x
First, according to the characteristics of UAV sensor data, a new loss
field.linear_acceleration.y function is designed by introducing the NCC into the MSE loss function.
field.linear_acceleration.z NMSE allows the model to better fit the clean data.
field.magnetic_field.xfield.magnetic_field.yfield. Second, a multi-modal DAE is designed to extract multi-modal fea-
magnetic_field.z
tures using heterogeneous neural networks to improve the denoising
Features of attitude classes in field.commanded.Pitch
datasets field.measured.Pitch capability of the model; the boldfaced squares in the diagram indicate
field.commanded.Roll the vectors that the current model pays increasing attention to, such as
field.measured.Rollfield.commanded.Yawfield. the attitude category (A) vector focused on by the MLP in the data.
measured.Yaw Finally, the L-MDAE model is pre-trained with noise data, and the
trained parameters are frozen to form a denoising model. A fully con-
represents the optimal solution and β represents the mapping from Φ to nected layer is added after the denoising model to obtain the anomaly
Ψ in the function space. detection model. The central nodes of the heterogeneous UAV swarm are
As shown in the communication links of the multi-UAV swarm in abstracted as clients in federated learning, and the GCS controlling the
Fig. 3, link noise exists in the communication among individual UAVs UAV swarm is abstracted as the server in federated learning. Anomaly
and between UAVs and GCSs. UAVs integrate various sensors such as detection of UAV swarms is achieved by leveraging the efficient
IMU and GPS. The above-mentioned communication links introduce parameter transmission of federated learning and features such as pri-
noise to these heterogeneous sensors. Notably, the noise characteristics vacy protection of the heterogeneous UAV swarm.
of different sensors exhibit multi-modal features. The following subsections will discuss each of these parts in detail.
As shown in Eq. (9), x(t)GPS , x(t)Mag and x(t)Baro refer to the UAV
sensor observation data from multi-sensors, which have different sam- 4.2. NMSE loss function in L-MDAE
pling frequencies and data recording modes, presenting multi-modal
characteristics. Communication interference among UAV swarms during missions
⎧ can introduce noise in the sensor logs, making it difficult for data-driven
⎪
⎪ x(t)GPS = s(t)GPS + Δn(t)GPS anomaly detection models to capture the internal features of the data,
⎪
⎪
⎪
⎨ x(t)Mag = s(t)Mag + Δn(t)Mag resulting in poor performance of these models.
(9) In the current study, we first add Gaussian random noise with mean μ
⎪
⎪ ⋯
⎪
⎪
⎪
⎩ and standard deviation ϑ to the original clean data according to Eq. (10)
x(t)Baro = s(t)Baro + Δn(t)Baro as a way to simulate the data noise generated by the UAV swarm flight.
This study categorises sensors into time-series recording sensors and Fig. 6 shows the change in data before and after adding noise to the clean
attitude sensors based on the characteristics of different sensors. These data.
two types of sensors exhibit heterogeneity and require different neural The rationale for using Gaussian noise in this study is that it more
networks for processing. As shown in Table 1, this study extracted some accurately simulates the unpredictable noise found in real-world envi-
sensor data features from the ALFA dataset to illustrate the above- ronments because in real environments, noise is often not caused by a
mentioned definitions for the ease of understanding. This study single source but by a complex of noises from many different sources.
We consider real noise as the sum of many random variables with

5
Y. Lu et al. Computers & Industrial Engineering 196 (2024) 110454

Fig. 4. UAV swarm anomaly detection model based on FL and L-MDAE.

6
Y. Lu et al. Computers & Industrial Engineering 196 (2024) 110454

Fig. 5. Gaussian noise density curves at different levels.

Fig. 6. Gaussian noise added to the original data.

different probability distributions and each random variable to be in- ( )

dependent. According to the Central Limit Theorem (Dudley, 2014); the (x− μ)2
1 (10)
−
2ϑ2
normalised sum of independent random variables will approach a f(x) = √̅̅̅̅̅̅ × exp
Gaussian distribution as the number of noise sources increases. Fig. 5 2πϑ
shows the Gaussian noise density curves under different noise levels. Based on UAV sensor data noise characteristics, this study introduces
The data distribution of Gaussian white noise exhibits certain differ- the NCC into the MSE loss function and proposes a loss function that can
ences across various levels of variance and mean. According to the re- better fit noisy data than MSE: NMSE; its calculation method is shown in
sults of previous studies (Zhao et al., 2019; Shao et al., 2020) and the
Eq. (11). In Eq. (11), XK refers to the data with added noise and Xʹ refers
characteristics of UAV channel noise, this study determines that the
to the vectors processed by L-MDAE.
noise is Gaussian noise with a mean value of 0.1 and a variance of 0.2.

7
Y. Lu et al. Computers & Industrial Engineering 196 (2024) 110454

Fig. 7. Example of NCC and loss functions.

⎧ ⎡ ⎛ ⎞⎤⎫
⎪
⎪ ∑n ( K ) ⎪
⎪
1∑n ( K i=1 Xi × Xi
)2 ⎨ ⎢ ⎜ ʹ ⎟⎥⎬
NMSE = X − Xʹ + 1− ⎢sigmoid⎜√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ ⎟ ⎥
̅ ⎠⎦ # (11)
n i=1 i i
⎪
⎪
⎣ ⎝ (∑
n K 2
) ( ∑n ( ʹ ) 2 ) ⎪
⎪
i=1 Xi
⎩ i=1 (Xi ) × ⎭

As shown in Eq. (11), the NMSE loss function consists of the MSE loss To ensure that the model accurately utilises the NCC for loss calcu-
function and the NCC. Adding the NCC to the loss function can more lation, the sigmoid function is used in this study to control the value of
effectively measure the waveform differences between noisy data and the NCC between 0 and 1, and sigmoid(NCC) ∈ [0, 1] and
data after denoising by the model compared with MSE, allowing the 1 − (sigmoid(NCC)) are used for loss inversion. As shown in Fig. 7, the
model to adjust parameter weights based on the loss and improve the NCC value will be extremely small when there is significant waveform
fitting effect of the model on clean data. Notably, when NCC ∈ [ − 1, 1], dissimilarity between the input and reconstructed vectors. This obser-
the larger its value, the more similar the waveforms of the two signals; vation suggests that the ability of the model to fit clean data is currently
conversely, the lower its value, the lower the similarity between two inadequate, and adjusting the weight parameters of the model would
waveforms. require an increase in the loss of the model. Therefore, this study uses

8
Y. Lu et al. Computers & Industrial Engineering 196 (2024) 110454

Fig. 8. Noise removal using L-MDAE.

loss inversion with 1 − (sigmoid(NCC)). By contrast, consider the second convolutional operation, ⊗ refers to the convolutional operation and bi
example in Fig. 7; when the waveform of the reconstructed vector is refers to the bias of the convolutional layer.
similar to the input vector, the NCC value becomes very high. This in- ( )
∑M
dicates that the model already possesses a strong capability to fit clean yiEncoderʹ
= ReLU j
x ⨂ωi + bi
K
(12)
j=1 i
data, and consequently, the loss value of the model should decrease at
this stage.
yEncoderʹ = MaxPool( yEncoderʹ, w) (13)
i,p i

4.3. Multi-modal noise–removal module in L-MDAE ( )

fθEncoder cnn
= ReLU ωi,f• yEncoder
i,p
ʹ+b
i,f (14)
Traditional DAEs lack the ability to utilise heterogeneous neural
The obtained feature vector is pooled using Eq. (13) to reduce the
networks for extracting multi-modal features from multi-modal sensors,
dimensionality of the convolutional layer, and further feature extraction
limiting the denoising capability of the model. The proposed L-MDAE
is performed. In Eq. (13), yEncoder refers to the output of the pooling layer
ʹ
model consists of two parts of heterogeneous neural networks: first, a i,p

one-dimensional CNN, which is a type of convolution that can be used and w refers to the window size. Higher-level features are extracted
for processing sequential data, enables L-MDAE to obtain more crucial using Eq. (14), where fθEncoder cnn refers to the final output of the encoder
features about UAV failures. During feature extraction, the CNN pays part of the CNN after multiple convolutional pooling, fθEncoder cnn ∈ RL×N ,
increasing attention to time-based features, such as the bolded T vector L refers to the number of input samples and N refers to the dimension of
in Fig. 8. the input sample features. ωi.f , bi,f refer to the corresponding weight
Next is the MLP, which has a strong nonlinear feature extraction matrix and bias of the fully connected layer, respectively.
capability due to its multilayer structural characteristics. This enables Herein, MLP is used in parallel for nonlinear feature extraction after
the L-MDAE to obtain more nonlinear features about the UAV and pay 1D CNN. Nonlinear feature extraction is calculated as shown in Eq. (15),
increasing attention to features based on the UAV’s attitude during where N refers to the dimensions of the data, ωi , b refer to the weights
feature extraction, such as the bolded A vector in Fig. 8. The L-MDAE and biases, respectively, and fθEncoder mlp
refers to the output vectors after
model proposed in this study integrates the features extracted by mul-
being mapped by the MLP, fθEncoder mlp
∈ RL×N . L refers to the number of
tiple modal neural networks, which gives the model stronger denoising
input samples, and N refers to the dimension of the input sample
and anomaly detection abilities.
features.
( )
4.3.1. Encoder in the L-MDAE model Encoder mlp
∑N
Fig. 8 illustrates the L-MDAE model. In the encoder part, local fea- fθ = ReLU i=1
ωi • xi + b
K
(15)
tures in UAV fault data are extracted first through 1D CNN. We consider
{ } {( ) ( K K)
the input vector as XK = X1K , X2K , X3K ...XNK = xK 1 , y1 , x2 , y2 , ...
K ( )
( K K ) ( K K )} ΦEncoder = Concatenate fθEncoder cnn
, fθEncoder mlp
(16)
xi , yi , ... xN , yN and N as the number of samples. The 1D CNN
operation can be performed using Eq. (12). In Eq. (12), yEncoder refers to
ʹ
i Then, this study uses the concatenate method to fuse the output
the output of the convolutional layer of the encoder, xK i refers to the
vectors of the 1D CNN and the MLP, enabling the model to obtain fea-
j
vector of inputs, ωi refers to the convolutional kernel used to perform the tures from heterogeneous neural networks. Concatenate can splice two

9
Y. Lu et al. Computers & Industrial Engineering 196 (2024) 110454

Fig. 9. FL anomaly detection model for UAV swarm.

vectors together. In Eq. (16), ΦEncoder refers to the vector after fusing the ΦEncoder is decoded in parallel in the decoder part using MLP, which is
output of the CNN and MLP, ΦEncoder ∈ R2L×N . L refers to the number of calculated as shown in Eq. (20). Further, the decoded vectors are fused
input samples, and N refers to the dimension of the input sample together using the concatenate method, as shown in Eq. (21); thus, the
features. model obtains the decoded vectors from the multimodality. ΦDecoder is the
feature vector after fusing the output of the multi-modal neural network,
4.3.2. Decoder in the L-MDAE model ΦDecoder ∈ R2L×N . L refers to the number of input samples, and N refers to
In the current study, ΦEncoder is decoded using the decoder L-MDAE to the dimension of the input sample features. After encoder and decoder,
obtain a higher-order vector representation. In Eq. (17), ΦiEncoder refers to the reconstructed vector processed by L-MDAE is obtained.
the vector encoded by the L-MDAE encoder and yDecoder refers to the
ʹ ( )
i ∑N
Encodermlp
output of the convolutional layer of the decoder. The output of the fθ = ReLU i=1
ωi • ΦEncoder + b
i
(20)
convolutional layer is downscaled using the maximum pooling method
of Eq. (18), and yDecoder refers to the output of the convolutional layer.
ʹ
i
( )
Decodermlp
Further feature extraction is performed on the pooled vectors using the ΦDecoder = Concatenate fθDecodercnn , fθ (21)
fully connected network in Eq. (19). In Eq. (19), fθDecoder cnn refers to the ∑N ( )
final output of the decoder part of the CNN after multiple convolutional i=1 NMSE xK
i , Xi
* * * *
ʹ
θen en de de
mlp , θcnn , θmlp , θcnn = arg min (22)
pooling, fθDecoder cnn ∈ RL×N , L refers to the number of input samples and *
θen
* * *
,θen ,θde ,θde n
mlp cnn mlp cnn
N refers to the dimension of the input sample features.ωi.f , bi,f corre-
spond to the weight matrix and bias of the fully connected layer, L-MDAE employs the objective function presented in Eq. (22) to
* * * *
respectively. optimise its parameter matrix. In Eq. (22), θen en de de
mlp , θcnn , θmlp and θcnn refer to
( ) the parameter matrices of encoder and decoder in L-MDAE and xK
i and xi
ʹ
∑M
yi
Decoderʹ
= ReLU Φ i j
⨂ωi + bi (17) refer to the vectors after encoding and decoding by L-MDAE.
j=1 Encoder

( ) 4.4. Anomaly detection using federated learning and L-MDAE

yDecoder = MaxPool yDecoder ,w (18)
ʹ ʹ
i,p i

( ) Owing to the limited computing resources of the central node in the

fθDecodercnn = ReLU ωi,f• yDecoder + bi,f (19) UAV swarm, this study pre-trains L-MDAE and then freezes its param-
ʹ
i,p
eters to use it as a noise-removal model, thereby reducing the number of

10
Y. Lu et al. Computers & Industrial Engineering 196 (2024) 110454

Fig. 10. Example for L-MDAE.

11
Y. Lu et al. Computers & Industrial Engineering 196 (2024) 110454

training parameters required for the anomaly detection model. The Table 2
operational description of the UAV swarm anomaly detection model Fault distribution of the ALFA dataset.
based on FL and L-MDAE proposed in this study is shown in Algorithm 1. Faulty part Type of failure Frequency
The loss of federated learning is shown in Eq. (24), where Loss(w)
Engine Full power loss 23
refers to the global loss, Lossi (w) refers to the loss of the ith UAV control Rudder Stuck to left 1
( )
node and fi w; XijK ; YijK refers to a client training example. Stuck to right 2
Elevator Stuck at zero 2
Algorithm 1 Anomaly detection using federated learning and L-MDAE Left aileron Stuck at zero 3
* * * * Right aileron Stuck at zero 4
Input: θen en de de
mlp , θcnn , θmlp , θcnn is the parameter for encoder and decoder in L-MDAE; Θ is a Both ailerons Stuck at zero 1
parameter of the fully connected layer after L-MDAE; K is the number of UAV Rudder and aileron Stuck at zero 1
control nodes; E is the number of global rounds; training set XK . No fault No fault 10
Total 47
* * * *
Output: θen en de de
mlp , θcnn , θmlp , θcnn and Θ.
* * * *
Initialisation: Freeze the parameters (θen en de de
mlp , θcnn , θmlp and θcnn ) of the pre-trained L-
MDAE and add a fully connected layer behind it (Θ).
1 FOR each round r = 0,1, …, E − 1 DO: conducted on these datasets to demonstrate the reliability of the L-
2 m←MAX(C × K, 1) MDAE model from various perspectives. The following subsections
3 Sr ←Randomly select m clients
( K )
present experiment details and dataset introductions. This study uses the
4 XKr = RandomlyDivide X , Sr configuration parameters provided in Table 3 to train the L-MDAE
model.
* * * *
5 ▹ Send θen en de de
mlp , θcnn , θmlp , θcnn and Θr to selected k ∈ Sr
6 FOR each client k ∈ Sr in parallel DO: As there is no specific dataset for training UAV swarm anomaly
( )
7 Θkr+1 ←ClientTraining Θr , Xkr detection models, this study divides the ALFA, TLM, ITS_Cyber,
8 END FOR
1 ∑
ITS_Physical and N_BaIoT datasets into equal parts according to the
9 Θr+1 = × k∈Sr Θkr+1 number of UAV control nodes to simulate the abnormal data of the UAV
K
10 END FOR swarm.

1 ∑N− 1 K
Trainloss = − y log(̂
yn) (23) 5.1. Dataset description, pre-processing, evaluation metrics and model
N n=0 n
details
∑n ∑n 1 ∑ni ( )
Loss(w) = Lossi (w) = i=1 n
fi w; XijK ; YijK (24)
i=1
i
j=1 The ALFA (Keipour et al., 2021) dataset is collected using the
Fig. 9 demonstrates the system architecture of the UAV swarm Carbon-Z T-28 UAV platform. This dataset includes 47 instances of UAV
anomaly detection model proposed in this study. The model consists of flight logs in normal or abnormal states. These logs contain raw flight
three steps: data and CSV files processed by the author of (Keipour et al., 2021). The
In the first step, the server reads the parameters of the trained L- 47 flight instances comprise 23 engine failure scenarios and 24 failure
MDAE model and adds the fully connected layer behind it. The model scenarios for seven other major components, e.g. engines and elevators.
divides the dataset among k clients (UAV central nodes) to simulate the The total flight time under normal conditions is 66 min, whereas the
training of the client’s parameters. The model initialises both the frozen flight time after failure is 13 min. Furthermore, the author of (Keipour
parameters and the parameters of the fully connected layer and then et al., 2021) has provided abnormal labels for evaluating anomaly
transmits them to the central node of the UAV swarm. The first step detection models. The breakdown of anomalies is shown in Table 2.
corresponds to lines 1–5 of Algorithm 1. In this study, we use the method proposed in literature (Deng et al.,
In the second step, the central node of each sub-swarm receives the 2024) to process the ALFA dataset, and the label distribution of the
parameters from the server (GCS) and trains the local anomaly detection processed dataset is shown in Fig. 11(b).
model using the cross-entropy loss function shown in Eq. (23) with the The TLM dataset (Yang et al., 2023) is a UAV failure dataset based on
simulated data of this swarm; the parameters are uploaded to the server a software-in-the-loop simulation environment. This environment sim-
after training. This corresponds to lines 6–8 of Algorithm 1. ulates common UAV failures by varying the internal physical parameters
In the third step, the server uses the average aggregation algorithm to of the UAV. The TLM dataset uses a quadrotor UAV as the data collection
unify the parameters from each client (the control node of the UAV) and platform, simulating common UAV anomalies such as engine failures,
distribute them to the central nodes of the sub-swarms. Step 3 corre- remote controller malfunctions and accelerometer faults during the UAV
sponds to line 9 of Algorithm 1. flight. The research team for this study simulated the above-mentioned
Steps 1–3 are repeated until the model reaches the maximum number faults and achieved raw data collection and processing. The TLM dataset
of iterations to complete the anomaly detection of the UAV swarm. is processed using the method proposed in a previous study (Deng et al.,
Fig. 10 shows a simple example of the L-MDAE model. First, noise is 2024); and the label distribution of the processed dataset is shown in
added to the original data to obtain XK ; the L-MDAE is trained using the Fig. 11(a).
NMSE loss function and the model needs to minimise the reconstruction The data collection platform for ITS_Cyber (Hassler et al., 2023) and
ITS_Physical (Hassler et al., 2023) is the DJI Tello EDU UAV. The ITS_-
loss of XK and Xʹ. Finally, after the converged L-MDAE denoising model
Cyber and ITS_Physical dataset covers a range of attacks, such as denial
is trained, its parameters are frozen, and the fully connected layer is
of service, replay, evil twin and false data injection attacks, and includes
added later. The training of the UAV swarm anomaly detection model is
cyber state information with dynamic physical data of UAVs under the
completed with federated learning and anomaly detection. The intro-
above-mentioned attacks. In this study, the data of UAVs under DoS
duction of the dataset used in this study and its pre-processing process
attack and replay attack are extracted, and the label distribution of the
will be introduced in detail in Section 5.
processed dataset is shown in Fig. 11(d) and (e).
N_BaIoT (Meidan et al., 2018) includes network traffic extracted
5. Experimentation
from nine real IoT devices, including normal traffic, as well as five types
of attack traffic from the Gafgyt and Mirai botnet families. The content
This study validates the performance of the L-MDAE model on ALFA,
of the dataset includes the statistical characteristics of packet-to-packet
TLM, ITS_Cyber, ITS_Physical and N_BaIoT datasets. Basic performance,
traffic, 10 types of attacks and normal traffic. In this study, we process
ablation and comparative experiments with other relevant models were
the dataset using the data processing method from literature (Chen

12
Y. Lu et al. Computers & Industrial Engineering 196 (2024) 110454

Fig. 11. Distribution of labels for processed datasets.

et al., 2023), and its labelling distribution is shown in Fig. 11(c).

TP + TN
Because UAV data mainly come from different sources or measuring Accuracy = (26)
TP + TN + FP + FN
tools, the data often have issues with different dimensions, which can
affect model training. Therefore, this study uses standardisation func- TP
tions to process the dataset to reduce the impact of data on the model. Precision = (27)
TP + FP
The standardisation function is calculated as shown in Eq. (25), where Z
refers to the processed features, X refers to the training samples to be TP
Recall = (28)
processed and μ and σ are the mean and standard deviation of the TP + FN
training samples, respectively.
2 × Precision × Recall
X− μ F1 = (29)
Z= (25) Precision + Recall
σ
This study assesses the proposed L-MDAE model using indicators FP
FPR = (30)
from Eqs. (26)–(31). In these equations, TP represents positive samples FP + TN
predicted as positive, FP corresponds to negative samples predicted as TP
positive, TN denotes negative samples predicted as negative and FN TPR = (31)
TP + FN
represents positive samples predicted as negative by the model.

13
Y. Lu et al. Computers & Industrial Engineering 196 (2024) 110454

Table 3 Table 5
Detailed structure of L-MDAE. Basic performance of L-MDAE under centralised and swarm scenarios.
Layer Description Configuration Datasets Experimental Test Precision F1 Recall
scenario Accuracy
Input layer Input layer in L-MDAE Shape = (M)
Dense layer MLP of encoder Shape = (512), Activation = ALFA Centralised 0.9926 0.9891 0.9926 0.9847
ReLU FL-Distributed 0.9901 0.9876 0.9900 0.9849
Dense layer MLP of encoder Shape = (576), Activation = TLM Centralised 0.9886 0.9852 0.9886 0.9760
ReLU FL-Distributed 0.9849 0.9619 0.9840 0.9792
Reshape layer Reshape the input Shape = (N,1) ITS-Cyber Centralised 0.8272 0.8306 0.8272 0.8380
Conv1D layer Convolutional layer of Filters = 4, Kernel_size = 6, FL-Distributed 0.8142 0.8243 0.8119 0.8265
encoder Activation = ReLU ITS- Centralised 0.9923 0.9867 0.9923 0.9876
MaxPool1D MaxPool layer of encoder Size = 2 Physical FL-Distributed 0.9903 0.9839 0.9903 0.9831
layer N-BaIoT Centralised 0.9809 0.8718 0.9809 0.8706
Flatten layer Flatten layer of encoder / FL-Distributed 0.9721 0.8207 0.9721 0.8652
Dense layer Calculate the convolved Shape = (576), Activation =
vector in encoder ReLU
Concatenate Splice the computed vector /
layer
Dense layer Hidden layer Shape = (128), Activation = Table 6
ReLU Performance of the L-MDAE model on the N-BaIoT dataset with different sample
Dense layer MLP of decoder Shape = (512), Activation = sizes.
ReLU
Dense layer MLP of decoder Shape = (576), Activation = N-BaIoT Experimental Test Precision F1 Recall
ReLU datasets scenario Accuracy
Dense layer Output of MLP in decoder Shape = (M), Activation = Original Centralised 0.9809 0.8718 0.9809 0.8706
ReLU dataset FL-Distributed 0.9721 0.8207 0.9721 0.8652
Reshape layer Reshape the vector Shape = (N,1) Scenario 1 Centralised 0.9740 0.9586 0.974 0.9582
Conv1D layer Convolutional layer of Filters = 4, Kernel_size = 6, FL-Distributed 0.9351 0.9036 0.9337 0.8899
decoder Activation = ReLU Scenario 2 Centralised 0.9810 0.9784 0.9810 0.9784
MaxPool1D MaxPool layer of decoder Size = 2 FL-Distributed 0.9449 0.9395 0.9436 0.9358
layer Scenario 3 Centralised 0.9798 0.9781 0.9798 0.9782
Flatten layer Flatten layer of decoder / FL-Distributed 0.9581 0.9553 0.9567 0.9541
Dense layer Calculate the convolved Shape = (576), Activation =
vector in decoder ReLU
Dense layer Output of CNN in decoder Shape = (M), Activation =
ReLU
Concatenate Splice the computed vector / Table 7
layer Different sample size scenarios of the N-BaIoT dataset.
Dense layer Output the vector calculated Shape = (M), Activation =
by L-MDAE ReLU Original dataset Scenario 1 Scenario 2 Scenario 3

Normal 27,962 27,962 27,962 27,962

Scan attack 3,589 3,589 10,500 14,000
Junk attack 3,542 3,542 10,500 14,000
Table 4 combo attack 3,472 3,472 10,500 14,000
Experimental parameters. Syn attack 2,126 3,500 10,500 14,000
UDP attack 2,079 3,500 10,500 14,000
Experimental environment Parameters
Scan_mirai attack 2,067 3,500 10,500 14,000
Num_UAV_Clients 10 Udpplian attack 2,067 3,500 10,500 14,000
Min_Fit_UAV_Clients 6 ACK attack 2,057 3,500 10,500 14,000
Min_Evaluate_UAV_Clients 5 TCP attack 207 3,500 10,500 14,000
Min_Available_UAV_Clients 7 UDP_gyfgyt_attack 197 3,500 10,500 14,000

The detailed structure used by the L-MDAE model proposed in this control nodes participating in the simulated swarm environment,
study is shown in Table 3. In this table, M refers to the feature dimension ‘Min_Fit_UAV_Clients’ refers to the minimum number of UAV control
of the input sample. nodes participating in training, ‘Min_Evaluate_UAV_Clients’ refers to the
minimum number of nodes participating in evaluating the model and
‘Min_Available_UAV_Clients’ refers to the minimum number of UAV
5.2. Basic performance experiment
control nodes that remain online. In Table 5, ‘FL-Distributed’ refers to
the experimental results of the UAV swarm anomaly detection model in
This study uses Flower1 as the federated learning simulation envi-
the swarm context.
ronment framework and TensorFlow2 as the deep learning framework to
Table 5 shows the performance of the proposed L-MDAE model in
conduct simulation experiments. Owing to the characteristics of feder-
different experimental scenarios. Experimental scenarios include cen-
ated learning, such as multi-party participation, distributed computing
tralised anomaly detection and UAV swarm–federated learning anomaly
and privacy protection, it can use limited computing resources to train
detection. In this study, experiments are conducted four times in the
and deploy efficient swarm anomaly detection models while ensuring
centralised scenario, and all the metrics of centralised training presented
the privacy of heterogeneous UAV swarms. Table 4 provides experi-
in Table 5 are the average of the four experiments. In this table, the
mental parameters such as the number of UAV control nodes and the
boldfaced data indicate the highest values of the experiment.
minimum number of nodes participating in the calculation under the
In Table 5, all the evaluation metrics for centralised training surpass
simulation environment.
those obtained from distributed swarm–federated learning across the
As shown in Table 4, ‘Num_UAV_Clients’ refers to the number of UAV
five datasets. This is because in centralised training, the model has ac-
cess to all the training data, which means that the model has access to a
1 larger, richer and more comprehensive dataset. This access contributes
https://flower.ai/.
2
https://www.tensorflow.org/.
to the improved accuracy of the model. However, in federated learning,

14
Y. Lu et al. Computers & Industrial Engineering 196 (2024) 110454

Fig. 12. ROC curves of the model with different datasets.

15
Y. Lu et al. Computers & Industrial Engineering 196 (2024) 110454

Fig. 13. Training loss and test accuracy of models in the context of swarm anomaly detection.

the data are distributed to different clients, resulting in different devices The convergence of the model during federated learning distributed
having different data, making the model less effective than centralised training for swarm anomaly detection is illustrated in Fig. 13. Around
training. the 10th iteration, the training loss of the model converged on these five
Table 5 shows that the model performs well on the ALFA, TLM and datasets. Subsequently, the aggregated global parameters after model
ITS-Physical datasets but does not perform as well on the ITS_Cyber and iterations are used as the global model, and the performance of this
N-BaIoT datasets. The ITS_Cyber dataset contains network data from model is tested using the test set. As shown in Fig. 13, the model has
UAVs subjected to DoS attacks, replay attacks and normal conditions. already achieved convergence in testing accuracy by the 5th iteration. In
Unlike the other datasets, the characteristics of the DoS attack data in conclusion, the proposed L-MDAE model demonstrates strong conver-
the ITS_Cyber dataset are similar to those of the normal data, making it gence, requiring only a small number of iterations to converge, making it
difficult for the model to distinguish between the two labels. suitable for dynamic anomaly detection in UAV swarms.
In addition, the model exhibits lower precision and recall on the N-
BaIoT dataset owing to the extremely unbalanced sample labelling. As 5.3. Ablation experiments
shown in Fig. 11(c), there are only 207 records for TCP_Attack and 197
records for UDP_gyfgyt_Attack in the dataset, which account for 0.42 % To verify the usefulness of the NMSE loss function and the multi-
and 0.4 % of the total samples, respectively. This imbalance makes it modal denoising module proposed in this study, the above compo-
difficult for the model to learn the feature distribution of these labels. nents are removed and the comprehensive performance of the model is
To demonstrate the effectiveness of the proposed model, this study trained and tested centrally.
uses the SMOTE (Chawla et al., 2002) algorithm to populate the un- The experimental results are shown in Table 8; the boldfaced data
balanced samples and further expands the number of populated samples indicate the highest values of the experiment. In Table 8, MDAE refers to
to explore the variation in model performance. The experimental results the experimental results using the MSE loss function after excluding the
are shown in Table 6, and the number of samples included in the NMSE loss function; DAE refers to the DAE with the MSE loss function
experimental scenarios in Table 6 is shown in Table 7. after excluding the multi-modal denoising module and NMSE loss
Tables 6 and 7 show that with an increase in the number of samples function. Remarkably, the experimental results for the N-BaIoT dataset
generated by the SMOTE algorithm, precision and recall also exhibit an in Table 8 are based on data from scenario 2.
upward trend. However, the lower quality of the generated data The experimental results in Table 8 demonstrate that the NMSE loss
compared to real data leads to a decrease in accuracy and F1 score. This function and the multi-modal denoising module significantly impact
drawback is further amplified, especially in distributed federated noise-removal and model performances across the five datasets. After
learning for swarm anomaly detection. For example, in scenario 1, the removing these modules, the model experiences varying degrees of
accuracy of the distributed federated learning for the swarm anomaly decline in accuracy, precision, F1 score and recall. In the TLM dataset,
detection model is 0.9351 and the precision is 0.9036. by replacing the NMSE loss function with the MSE loss function, the
Figs. 12 and 14 show the ROC curves and confusion matrices of the precision value of the model dropped by ~ 1 %. In addition, the model
models on different datasets in centralised training mode, and their sub- precision decreased by approximately 2 % after removing the multi-
figures represent the ROC curves and confusion matrices under different modal denoising module. This is particularly evident in the ITS_Cyber
datasets. Fig. 12(a), (b), (d) and (e) shows that the model performs well dataset. When the MSE loss function was used instead of the NMSE loss
on all four datasets, with ROC values close to 1. Fig. 12(c) shows that the function, the accuracy value of the model decreased by ~ 1 % and
performance of the model on the ITS_Cyber dataset is not as good as the precision decreased by ~ 2 %. After removing the multi-modal denois-
other four datasets, with an ROC value of approximately 0.94. Similarly, ing module, the model experienced a decline of approximately 5 % in
Fig. 14(a), (b), (d) and (e) shows that the model performs well on the accuracy, precision and F1 values. As shown in Fig. 15, for a more visual
above four datasets. Fig. 14(c) shows that the performance of the model presentation, this study uses bar charts to show the test accuracy of the
on the ITS_Cyber dataset is not as good as the other four datasets. This model on different datasets as the modules are removed.
confusion matrix visually shows that the model can effectively classify
replay attacks but still struggles with fully classifying DoS attacks. This is 5.4. Comparison with other relevant models
because the ITS_Cyber dataset has a high similarity between DoS attack
samples and normal samples, making it challenging for the model to In the current study, we use MCNN-LSTM (Chen et al., 2021), PB-DID
identify them completely. (Zeeshan et al., 2021) and LCCDE (Yang et al., 2022:) to compare the

16
Y. Lu et al. Computers & Industrial Engineering 196 (2024) 110454

Fig. 14. Confusion matrix of the model with different datasets.

17
Y. Lu et al. Computers & Industrial Engineering 196 (2024) 110454

Table 8 and negative samples, with only 973 records for two abnormal labels
Performance of the L-MDAE model after removing relevant components. compared to 4,290 records for normal samples. Machine learning
Datasets Models Test Accuracy Precision F1 Recall models have strong classification capabilities with few sample data, but
this ability diminishes as the sample size increases.
ALFA L-MDAE 0.9926 0.9891 0.9926 0.9847
MDAE 0.9883 0.9867 0.9883 0.9761 To demonstrate this hypothesis, this study used the SMOTE algo-
DAE 0.971 0.9724 0.971 0.9600 rithm to increase the number of two classes of abnormal samples to
TLM L-MDAE 0.9886 0.9852 0.9886 0.9760 4,290 records. The experimental results are shown in Table 10. Table 10
MDAE 0.9842 0.9714 0.9842 0.9733 shows that as the sample size increases, the indicators of the proposed L-
DAE 0.9795 0.9655 0.9795 0.9604
ITS_Cyber L-MDAE 0.8272 0.8306 0.8272 0.8380
MDAE model show little difference compared to before. However, the
MDAE 0.8106 0.812 0.8106 0.8196 performance based on machine learning models is gradually decreasing,
DAE 0.7735 0.7806 0.7735 0.7889 which is consistent with the previous conjecture.
ITS_Physical L-MDAE 0.9923 0.9867 0.9923 0.9876 To further investigate the effect of sample size on L-MDAE and ma-
MDAE 0.9879 0.9780 0.9879 0.97565
chine learning–based models, this study conducted experiments on the
DAE 0.9791 0.9606 0.9791 0.9569
N_BaIoT L-MDAE 0.9810 0.9784 0.9810 0.9784 LCCDE model and L-MDAE model with different numbers of sample
MDAE 0.9733 0.9695 0.9733 0.9695 scenarios provided in Table 7 and the experimental results are shown in
DAE 0.8868 0.8722 0.8868 0.86965 Fig. 17. As shown in this figure, the performance of the machine lear-
ning–based models deteriorates as the number of samples increases as
well as the change in the accuracy of the L-MDAE model decreases and
performance with the L-MDAE model proposed in this study under the
its recall increases. However, the performance of the L-MDAE model is
same parameter magnitude. MCNN-LSTM is a fault detection model for
more stable. The reasons can be summed up as follows. First, the ma-
fault diagnosis, which consists of CNN and LSTM. The model uses raw
chine learning–based model has a poorer fitting ability to the data in
vibration signals as inputs and uses two CNNs with different kernel sizes
high-dimensional sample space. Second, due to the noisy nature of the
to automatically extract features of the signals at different frequencies
data, the machine learning–based model has a poorer anti-noise ability.
from the raw data. An LSTM is then used to identify the fault type based
on the learned features. PB-DID is an IoT intrusion detection model that
conducts thorough feature selection and data pre-processing before
Table 9
inputting data into the model. It utilises an LSTM-based deep learning
Comparison of performance of different models.
architecture to detect DoS and DDoS attacks. The LCCDE model is a
model for intrusion detection in self-driving vehicles and connected Datasets Models Test Accuracy Precision F1 Recall
vehicles that is constructed by identifying the best-performing ML model ALFA L-MDAE 0.9926 0.9891 0.9926 0.9847
for each class or type of attack among three machine learning algorithms FL 0.9901 0.9876 0.9900 0.9849
(XGBoost et al.) and is an intrusion detection algorithm based on the MCNN-LSTM 0.8512 0.8535 0.8512 0.6133
PB-DID 0.8187 0.8398 0.8187 0.5323
integration of multi-modal machine learning models. LCCDE 0.9353 0.9340 0.9323 0.9353
The experimental results are shown in Table 9, where FL refers to the TLM L-MDAE 0.9886 0.9852 0.9886 0.9760
federated learning distributed-training swarm anomaly detection FL 0.9849 0.9619 0.9840 0.9792
model. The boldfaced data in the table indicate the highest values of the MCNN-LSTM 0.9661 0.9634 0.9661 0.9447
PB-DID 0.8931 0.7391 0.8931 0.7348
experiment. From Table 9, it can be seen that the proposed L-MDAE
LCCDE 0.9540 0.9533 0.9540 0.9531
model, as well as the swarm anomaly detection model, have higher ITS_Cyber L-MDAE 0.8272 0.8306 0.8272 0.8380
performance on the ALFA, TLM, ITS_Cyber and N_BaIoT datasets in FL 0.8142 0.8243 0.8119 0.8265
various metrics than on relevant contrast models. The accuracy of the MCNN-LSTM 0.7231 0.7544 0.7231 0.7436
ALFA dataset is 0.9926, which is 10 % higher than that of MCNN-LSTM, PB-DID 0.6798 0.6975 0.6798 0.6962
LCCDE 0.7156 0.7161 0.7156 0.7156
18 % higher than that of PB-DID and 6 % higher than that of LCCDE. As ITS_Physical L-MDAE 0.9923 0.9867 0.9923 0.9876
shown in Fig. 16, for a more intuitive presentation, this study uses bar FL 0.9903 0.9839 0.9903 0.9831
charts to demonstrate the testing accuracy of the contrast models on MCNN-LSTM 0.9455 0.8938 0.9455 0.9152
different datasets. PB-DID 0.7884 0.7539 0.7884 0.5423
LCCDE 1* 1* 1* 1*
On the ITS_Physical dataset, the metrics for the LCCDE model are all
N_BaIoT L-MDAE 0.9810 0.9784 0.9810 0.9784
1*. This occurs because the LCCDE model is an ensemble of machine FL 0.9449 0.9395 0.9436 0.9358
learning models, giving it strong classification capabilities for imbal- MCNN-LSTM 0.9096 0.9057 0.9096 0.8970
anced data with small sample sizes. Fig. 11(e) shows that the PB-DID 0.6347 0.5659 0.6347 0.5829
ITS_Physical dataset exhibits an extreme imbalance between positive LCCDE 0.7122 0.6926 0.7122 0.6845

Fig. 15. Ablation experiments on different datasets.

18
Y. Lu et al. Computers & Industrial Engineering 196 (2024) 110454

Fig. 16. Comparative experiments with other models.

‘Centralised’ refers to the experimental results using the L-MDAE model

Table 10 and ‘FL-Distributed’ refers to the experimental results deploying L-
Comparative experiment after adding samples of the ITS_Physical dataset. MDAE in a federated learning environment. This study compares the
Models Test Accuracy Precision F1 Recall performance of some machine learning models such as AdaBoost, naive
L-MDAE 0.9918 0.9918 0.9918 0.9917 Bayes, Random Forest and Random Tree on the AWID (the experimental
FL 0.9933 0.9944 0.9945 0.9945
LCCDE 0.4090 0.2497 0.4090 0.3088
Table 11
AWID distribution after process.
The L-MDAE model has a strong clean data–fitting ability as well as
Label name No. of samples
high-dimensional data processing ability, and its comprehensive ability
is higher than the machine learning–based anomaly detection model. fragmentation 31,180
probe_response 31,180
beacon 28,000
6. Discussion evil_twin 30,583
authentication_request 31,180
6.1. Generalisation experiments deauthentication 31,341
amok 28,963
cafe_latte 30,800
This study uses the Aegean Wi-Fi Intrusion Dataset (AWID) to test the arp 411,580
performance of our model in the WIFI IoT environment. AWID (Kolias normal 31,180
et al., 2015) is a public dataset containing normal traffic along with
many types of attacks against IEEE 802.11 networks. The dataset
development team collected the original data by configuring a typical Table 12
network environment, launching attacks on the target using Kali and Experiment result using the AWID.
capturing data packets. The experimental environment includes one
Model Test Accuracy Precision F1 Recall
desktop computer, two laptops, two smartphones, one tablet and one
smart TV. This study uses the method of literature (Chen et al., 2021) to Centralised 0.9993 0.9992 0.9993 0.9992
FL-Distributed 0.9989 0.9980 0.9989 0.9992
process the dataset. The data distribution in the processed dataset is AdaBoost 0.9028 0.9140 – 0.9030
shown in Table 11 below (the table stems from literature (Chen et al., Naive Bayes 0.9187 0.9230 – 0.9190
2021). Random Forest 0.9422 0.9470 – 0.9420
The experimental results are shown in Table 12. In Table 12, Random Tree 0.8927 0.9110 – 0.8930

Fig. 17. Performance comparison of machine learning–based model and L-MDAE with different sample sizes of the N-BaIoT dataset.

19
Y. Lu et al. Computers & Industrial Engineering 196 (2024) 110454

Fig. 18. Common time complexity.

results of the machine learning model stem from the literature (Chen
Table 13
et al., 2021). As can be seen from the table, the L-MDAE model proposed
Experimental environment configuration.
in this study has the highest accuracy and other experimental results are
better than the machine learning model. It can be seen from the above Experimental configurations Name

experimental results that the L-MDAE model has good multi-scenario CPU 11th Gen Intel(R) Core (TM) i7-11800H
expansion and adaptability and has good generalisation performance. GPU NVIDIA GeForce RTX 3050 Ti Laptop GPU
RAM 16 GB
Hard disk 512 GB
6.2. Complexity analysis Operating system Windows 11

As the L-MDAE model proposed in this study is mainly based on MLP

neural networks and CNNs, its main computing overhead stems from the Table 14
multiplication between matrices. To facilitate the analysis of the time Nomenclature.
complexity of the model, we assume that the input dimension is N, the Abbreviation Meaning
output dimension is M and the MLP has a total of L layers. Then, the time UAVs Unmanned aerial vehicles
complexity is expressed using Big O, as shown in the following Eqs. (32) GCS Ground control station
and (33): MSE Mean square error
( ) NCC Normalised correlation coefficient
∑L DAE Denoising autoencoder
O(MLP) = O l=1
(N × M) = O(N × M) (32) IMU Inertial measurement unit
GPS Global positioning system
MLP Multilayer perceptron
In the most extreme case, assuming that the input and output di-
( ) CNN Convolutional neural network
mensions are consistent, the time complexity is O N2 . LSTM Long short-term memory neural network
( )
∑L ( )
O(MLP) = O l=1
(N × N) = O N2 # (33) usually fixed). In the worst case, the number of input channels is equal to
the number of output channels and is represented by N. Then, the time
( )
Another fundamental component of L-MDAE is the CNN. Because of complexity of the CNN is O N2 .
its convolutional calculation characteristics, the main time consumption ( )
of this model arises from mapping the input channels to the output O(CNN 1) = O F2 × K2 × Cin × Cout (34)
channels. The time complexity of its single convolutional layer can be ( )
expressed using Eq. (34). Among these parameters, F represents the side ∑L
O(CNN) = O F2 × K2i × Ci− 1 × Ci (35)
length of the convolution kernel output feature map, K denotes the side i=1 i

length of the convolution kernel, Cin corresponds to the number of input

channels and Cout represents the number of output channels. The overall (
∑L
)
( )
time complexity of the CNN is shown in the Eq. (35), where L is the O(CNN) = O i=1
Ci− 1 × Ci = O N2 (36)
number of layers of the CNN. Eq. (36) is the simplified version of Eq.
(35). We consider F and K to be fixed constants (because the size of the The time complexity of the L-MDAE model is shown in Eq. (37). The
convolution kernel and the size of the input and output feature maps are

20
Y. Lu et al. Computers & Industrial Engineering 196 (2024) 110454

above-mentioned analysis shows that the time complexity of the L- Chen: Validation, Investigation. Rong Zeng: Validation.
MDAE model proposed in this study is s. The common time complexity
change function is shown in Fig. 18. The L-MDAE model has a low time Declaration of competing interest
complexity. Therefore, L-MDAE has strong practical significance.
( ) ( ) ( ) The authors declare that they have no known competing financial
O(LMDAE ) = O(MLP) + O(CNN) = O N2 + O N2 = O N2 (37)
interests or personal relationships that could have appeared to influence
the work reported in this paper.
6.3. Deployment details
Data availability
As the UAV control node is only responsible for collecting data, it
does not require excessive computing resources. The dynamic training Data will be made available on request.
process of the UAV swarm anomaly detection model is deployed on the
GCS. The GCS usually has strong computing resources; hence, the Funding statement & acknowledgments
training and deployment of the UAV swarm anomaly detection model
are considered feasible. This study lists the environment configuration This work was supported by the Innovation Team Funds of China
parameters used when conducting simulation experiments to provide a West Normal University under Grant No. KCXTD2022-3, and the Nan-
reference for the real deployment of the model in Table 13. chong City Science and Technology Plan Project under Grant No.
In addition, after completing the pre-training of the denoising model, 23YYJCYJ0142. Thank you to everyone who has contributed to this
the UAV swarm anomaly detection model contains only 549,449 train- work.
able parameters. This reduction in parameters contributes to improved
training speed while maintaining optimal performance, laying the Appendix A. Supplementary data
groundwork for the actual deployment of the model on UAV swarms.
Supplementary data to this article can be found online at https://doi.
7. Conclusions org/10.1016/j.cie.2024.110454.

The link noise generated by UAV swarms during operation creates References
multi-modal noise for multi-source UAV sensors, limiting the denoising
capability of data-driven swarm-based anomaly detection models and Ahn, H., & Chung, S. (2024). Deep learning-based anomaly detection for individual UAV
deteriorating their performance. In addition, the lack of data sharing vehicles performing swarm missions[J]. Expert Systems with Applications, 244, Article
122869.
involved in heterogeneous UAV swarms from multiple agencies makes it Air Accidents Investigation Branch reports, “AAIB investigation to DJI M600 Pro (UAS,
impossible for them to centrally train a single high-performance registration n/a) 131219”, Website 06 2020. [Online], Available: https://www.gov.
anomaly detection model, which further affects the stability of UAV uk/aaib-reports/aaib-investigation-to-dji-m600-pro-uas-registration-n-a-131219.
Al Amir, N., Marar, A., & Saeed, M. (2018) Eye in the sky: how the rise of UAVs will
swarms. This study proposes a UAV swarm anomaly detection model transfrom the oil & gas industry. In: Abu Dhabi international petroleum exhibition and
based on a multi-modal denoising autoencoder with federated learning conference. OnePetro.
(L-MDAE). First, perturbations are added to UAV anomalous sensor data Albiero, D., Garcia, A. P., Umezu, C. K., et al. (2022). Swarm robots in mechanized
agricultural operations: A review about challenges for research[J]. Computers and
to simulate the noise caused by the sensor data during UAV swarming Electronics in Agriculture, 193, Article 106608.
operations. Second, L-MDAE designs a new loss function (NMSE) ac- Alos, A. M., Dahrouj, Z., & Dakkak, M. (2020). A novel technique to assess UAV behavior
cording to the characteristics of UAV data noise by introducing the NCC using PCA-based anomaly detection algorithm[J]. International Journal of Mechanical
Engineering and Robotics Research, 9, 721–726.
into the MSE loss function, allowing the model to better fit the clean Altinors, A., Yol, F., & Yaman, O. (2021). A sound based method for fault detection with
data. Furthermore, L-MDAE uses multiple heterogeneous neural net- statistical feature extraction in UAV motors[J]. Applied Acoustics, 183, Article
works for multi-modal noise feature extraction so that the model can be 108325.
Asadzadeh, S., de Oliveira, W. J., & de Souza Filho, C. R. (2022). UAV-based remote
denoised from a multi-modal perspective, which improves the denoising
sensing for the petroleum industry and environ-mental monitoring: State-of-the-art
capability of the model. Finally, L-MDAE is pre-trained, and the trained and perspectives. Journal of Petroleum Science and Engineering, 208, Article 109633.
parameters are frozen to form a noise reduction model, after which a Bell, V., Rengasamy, D., Rothwell, B., et al. (2022). Anomaly detection for unmanned
fully connected layer is added to obtain an anomaly detection model. aerial vehicle sensor data using a stacked recurrent autoencoder method with
dynamic thresholding[J]. arXiv preprint arXiv:2203.04734.
The L-MDAE model is deployed on the client side of federated learning, Bronz, M., Baskaya, E., Delahaye, D., et al. (2020:). Real-time fault detection on small
which takes advantage of multi-party participation, distributed fixed-wing UAVs using machine learning[C]//2020 AIAA/IEEE 39th Digital
computation and efficient parameter transfer of federated learning to Avionics Systems Conference (DASC). IEEE, 1–10.
Brundage, M., Avin, S., Clark, J., et al. (2018). The malicious use of artificial intelligence:
perform anomaly detection in UAV swarms. The basic performance Forecasting, prevention, and mitigation[J]. arXiv preprint arXiv:1802.07228.
experiment proves that the L-MDAE model exhibits substantial Chawla, N. V., Bowyer, K. W., Hall, L. O., et al. (2002). SMOTE: Synthetic minority over-
improvement compared to the baseline model. The ablation experiment sampling technique[J]. Journal of artificial Intelligence Research, 16, 321–357.
Chen, J., Yang, T., He, B., et al. (2021). An analysis and research on wireless network
proves that the modules of the L-MDAE model play an essential role in security dataset[C]//2021 International Conference on Big Data Analysis and
denoising. The comparison experiment proves that the L-MDAE model Computer Science (BDACS). IEEE, pp. 80–83.
has stronger denoising ability than other models, and it is more suitable Chen, Y., Wang, J., Yang, T., et al. (2023). An enhancement method in few-shot scenarios
for intrusion detection in smart home environments[J]. Electronics, 12(15), 3304.
for anomaly detection in UAV swarms. However, the L-MDAE model still
Chen, X., Zhang, B., & Gao, D. (2021). Bearing fault diagnosis base on multi-scale CNN
needs to be improved regarding the classification ability of similar and LSTM model[J]. Journal of Intelligent Manufacturing, 32(4), 971–987.
samples. In the future, studies should focus on improving similar sample Deng, H., Lu, Y., Yang, T., et al. (2024). Unmanned Aerial Vehicles anomaly detection
model based on sensor information fusion and hybrid multimodal neural network
classification capabilities and the actual deployment capabilities of the
[J]. Engineering Applications of Artificial Intelligence, 132, Article 107961.
model through continued model simplification. Dudley, R. M. (2014). Uniform central limit theorems[M]. Cambridge University Press.
Hassler, S. C., Mughal, U. A., & Ismail, M. (2023). Cyber-physical intrusion detection
system for unmanned aerial vehicles[J]. IEEE Transactions on Intelligent
CRediT authorship contribution statement
Transportation Systems.
Hayawi, K., Anwar, Z., Malik, A. W., et al. (2023). Airborne computing: A toolkit for
Yu Lu: Writing – review & editing, Writing – original draft, Meth- UAV-assisted federated computing for sustainable smart cities[J]. IEEE Internet of
odology, Funding acquisition, Data curation, Conceptualization. Tao Things Journal.
He, X., Chen, Q., Tang, L., et al. (2022). Cgan-based collaborative intrusion detection for
Yang: Writing – review & editing, Supervision, Project administration, uav networks: A blockchain-empowered distributed federated learning approach[J].
Funding acquisition, Conceptualization. Chong Zhao: Resources. Wen IEEE Internet of Things Journal, 10(1), 120–132.

21
Y. Lu et al. Computers & Industrial Engineering 196 (2024) 110454

Horyna, J., Baca, T., Walter, V., et al. (2023). Decentralized swarms of unmanned aerial neural network and thermal images[J]. IEEE Transactions on Industrial Informatics, 17
vehicles for search and rescue operations without explicit communication[J]. (5), 3488–3496.
Autonomous Robots, 47(1), 77–93. South China Morning Post. HK$1 million in damage caused by GPS jamming that caused
Johnsen, S. O., Bakken, T., Transeth, A. A., Holmstrøm, S., Merz, M., Grøtli, E. I., & 46 UAVs to plummet during Hong Kong show. 2018, URL https://www.scmp.com/
Storvold, R. (2020). Safety and security of UAVs in the oil and gas industry. In news/hong-kong/law-and-crime/article/2170669/hk13-million-damage-caused-
Proceedings of the 30th European safety and reliability conference and the 15th gps-jamming-caused-46-UAVs.
probabilistic safety assessment and management conference, ESREL2020-PSAM15 Subbarayalu, V., & Vensuslaus, M. A. (2023). An intrusion detection system for UAV
organizers, Singapore. swarming utilizing timed probabilistic automata[J]. UAVs, 7(4), 248.
Keipour, A., Mousaei, M., & Scherer, S. (2021). Alfa: A dataset for uav fault and anomaly Vincent, P., Larochelle, H., Bengio, Y., et al. (2008). Extracting and composing robust
detection[J]. The International Journal of Robotics Research, 40(2–3), 515–520. features with denoising autoencoders[C]//Proceedings of the 25th international
Kolias, C., Kambourakis, G., Stavrou, A., et al. (2015). Intrusion detection in 802.11 conference on Machine learning, pp. 1096–1103.
networks: Empirical evaluation of threats and a public dataset[J]. IEEE Wang, Y., Wang, D., & Wang, J. (2015). A data driven approach for detection and
Communications Surveys & Tutorials, 18(1), 184–208. isolation of anomalies in a group of UAVs[J]. Chinese Journal of Aeronautics, 28(1),
Krishna, C. L., & Murphy, R. R. (2017). A review on cybersecurity vulnerabilities for 206–213.
unmanned aerial vehicles. In 2017 IEEE international symposium on safety, security and Yang, L., Li, S. B., Li, C. J., et al. (2023). Data-driven unsupervised anomaly detection and
rescue robotics (SSRR) (pp. 194–199). IEEE. recovery of unmanned aerial vehicle flight data based on spatiotemporal correlation
Meidan, Y., Bohadana, M., Mathov, Y., et al. (2018). N-baiot—network-based detection [J]. Science China Technological Sciences, 66(5), 1304–1316.
of IoT botnet attacks using deep autoencoders[J]. IEEE Pervasive Computing, 17(3), Yang, L., Li, S., Zhu, C., et al. (2024). Spatio-temporal correlation-based multiple
12–22. regression for anomaly detection and recovery of unmanned aerial vehicle flight
Motlagh, N. H., Kortoçi, P., Su, X., et al. (2023). Unmanned aerial vehicles for air data[J]. Advanced Engineering Informatics, 60, Article 102440.
pollution monitoring: A survey[J]. IEEE Internet of Things Journal. Yang, T., Lu, Y., Deng, H., et al. (2023). Acquisition and processing of UAV fault data
《Replicating peregrine attack strategies could help down rogue UAVs》https://www. based on time line modeling method[J]. Applied Sciences, 13(7), 4301.
ox.ac.uk/news/2017-12-04-replicating-peregrine-attack-strategies-could-help- Yang, L., Shami, A., Stevens, G., et al. (2022:). LCCDE: A decision-based ensemble
down-rogue-UAVs. framework for intrusion detection in the internet of vehicles[C]//GLOBECOM
Seo, S.-H., Lee, B.-H., Im, S.-H., & Jee, G.-I. (2015). Effect of spoofing on unmanned aerial 2022–2022. IEEE Global Communications Conference. IEEE, 3545–3550.
vehicle using counterfeited gps signal. Journal of Positioning Navigation and Timing, Zeeshan, M., Riaz, Q., Bilal, M. A., et al. (2021). Protocol-based deep intrusion detection
06, 57–65. for dos and ddos attacks using unsw-nb15 and bot-iot data-sets[J]. IEEE Access, 10,
Shao, H., Xia, M., Han, G., et al. (2020). Intelligent fault diagnosis of rotor-bearing 2269–2283.
system under varying working conditions with modified transfer convolutional Zhao, M., Zhong, S., Fu, X., et al. (2019). Deep residual shrinkage networks for fault
diagnosis[J]. IEEE Transactions on Industrial Informatics, 16(7), 4681–4690.

22