Robotics & Mechatronics Engineering

Federated Learning for Intrusion Detection System in IoT Devices

Submitted By:
Muhaiminul Islam
Syeda Faiza Ahmed

Submitted To:
Professor Dr. Lafifa Jamal
Department of Robotics & Mechatronics Engineering
University of Dhaka

12th July, 2023

TABLE OF CONTENTS
1 Introduction 1
1.1 Research Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2 Attacks on IoT 3
2.1 Botnet Attacks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 DDoS Attacks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.3 Firmware Exploitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.4 Data Privacy and Breaches . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.5 Credential Attacks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3 Intrusion Detection System 5

3.1 Importance of IDS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.2 IDS Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.3 Current Strategies and Limitations . . . . . . . . . . . . . . . . . . . . . . 8

4 Federated Learning 9
4.1 Federated Learning Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

5 Federated Learning for Intrusion Detection System 13

5.1 FL for Heterogeneous Anomaly Detection . . . . . . . . . . . . . . . . . . 13
5.2 FL for DDoS Attacks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
5.3 FL for privacy-preserving data handling . . . . . . . . . . . . . . . . . . . 15
5.4 FL for IDS in low-power IoT devices . . . . . . . . . . . . . . . . . . . . . 15
5.5 Challenges in Deploying FL in IDS . . . . . . . . . . . . . . . . . . . . . . 16

6 A Case Study: Federated IoT Attack Detection Using Decentralized Edge

Data [1] 18
6.1 Proposed Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
6.1.1 Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
6.1.2 Deep autoencoder design . . . . . . . . . . . . . . . . . . . . . . . 19
6.1.3 Security gateways . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
6.1.4 Multi-virtual workers process . . . . . . . . . . . . . . . . . . . . . 20
6.1.5 Federated learning simulation and implementation . . . . . . . . . 20
6.2 Research methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
6.2.1 Dataset & Performance metrics . . . . . . . . . . . . . . . . . . . . 21
6.2.2 Evaluation and testing goals . . . . . . . . . . . . . . . . . . . . . 22
6.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

7 Conclusion 23

1
8 24

2
1. Introduction

Intrusion Detection Systems (IDS) play a crucial role in securing networks by detecting
and responding to unauthorized activities or anomalies. In the context of IoT devices, IDS
becomes even more critical due to the large-scale deployment of interconnected devices with
diverse capabilities and vulnerabilities.

IoT networks pose unique challenges for intrusion detection due to the heterogeneous nature
of devices, resource limitations, and the distributed nature of the network. Traditional IDS
approaches designed for conventional networks may not be directly applicable or scalable to
IoT environments due to their inherent differences.

Federated Learning (FL) offers a promising solution for enhancing IDS in IoT networks. FL
leverages the distributed computing power of IoT devices and allows collaborative model
training without sharing raw data. By retaining data on the local devices and sharing only
model updates, FL ensures privacy while enabling the aggregation of knowledge from vari-
ous devices.

FL-based IDS in IoT can leverage the collective intelligence and local data from multiple
devices to enhance anomaly detection. Local models on IoT devices can learn from device-
specific data patterns and contribute to building a global model that represents a comprehen-
sive understanding of the network.

Furthermore, the decentralized and collaborative nature of FL in IDS allows for scalability
in IoT environments. Devices can autonomously detect and respond to intrusions, reducing
the reliance on a centralized server and enabling faster response times.

FL-based IDS in IoT can also mitigate the challenges of resource-constrained devices. In-
stead of relying on individual devices to perform resource-intensive computations, the col-
laborative training approach of FL allows devices to collectively contribute to model updates
without straining their limited resources.

Research in FL-based IDS for IoT is ongoing, with a focus on developing efficient algo-
rithms, improving model accuracy, scalability, and addressing the unique challenges posed
by IoT environments such as limited bandwidth, intermittent connectivity, and heterogeneous
device characteristics.

1
1.1. Research Questions

For our work, we want to investigate the existing solution for Intrusion Detection with Fed-
erated Learning. We also intend to study various IoT security aspects to develop a robust and
privacy-preserving IDS framework. For our exploration we intend to be guided through the
following research question:

1. What are the main challenges and future trends to be considered in the coming future
for the development of FL-enabled IDS for IoT scenarios?

2. What are available and standardized resources (dataset, evaluation protocols, etc. ) to
develop methodologies of intrusion detection systems for IoT devices?

3. How the analysis of existing FL-enabled IDS approaches for IoT based on a set of
identified criteria were performed?

After a comprehensive literature study, the paper [2] appeared to be a great match for ex-
ploring the First question where Federated Learning (FL)-enabled Intrusion Detection Sys-
tem (IDS) approach in the context of IoT, considering different data distributions and attack
types, and identifying challenges and future directions for FL-enabled IDS in real-world
scenarios was intensively studied. For investigating the last 2 questions, the literature [1] is
going to be considered where the authors proposed a federated-based approach using deep
autoencoders to detect botnet attacks in IoT devices by leveraging decentralized traffic data,
ensuring privacy and data security, and achieving high accuracy rates in anomaly detection
compared to centralized approaches.

2
2. Attacks on IoT

A deliberate attempt to take advantage of vulnerabilities in internet-connected devices in-

cluding smart homes, industrial control systems, and medical equipment is known as an
Internet of Things (IoT) assault. Attackers may take over the device, steal confidential data,
or utilise it as a component of a botnet for other illegal activities. As more devices are net-
worked, attacks on the Internet of Things (IoT) are becoming more frequent. Here are a few
illustrations of typical attack strategies that might harm IoT devices.

2.1. Botnet Attacks

A botnet is a network of compromised devices that is used to launch coordinated attacks.

These attacks are typically orchestrated by malicious actors who take over a huge number
of devices, frequently including IoT devices, and then employ them to carry out various
malicious activities[3].

Figure 2.1: Botnet Attacks on IoT.

3
Botnet Formation: A large number of devices must be infected with malware in order to
become part of a botnet, transforming them into zombies or automatons that can be com-
manded by the attacker.
Command and Control (C&C): When machines become harmed and infected, they link to
the attacker’s primary command and control system.
Malicious Activities:Attacks using botnets make advantage of infected devices to perform
malicious activities. These actions consist of conducting DDoS assaults, spamming, phish-
ing, stealing credentials, stealing data, and mining cryptocurrencies.[4]
Botnet Maintenance and Expansion: The ongoing administration and extension of a bot-
net network constitute botnet maintenance and expansion. In order to evade detection and
takedown attempts, this includes infecting new devices, upgrading malware on compromised
ones, and rotating the command and control infrastructure.

2.2. DDoS Attacks

DDoS (Distributed Denial-of-Service) attacks are deliberate efforts to prevent legitimate

users from accessing a system or network by flooding it with excessive traffic. These attacks
intentionally target IoT networks and devices, using inherent flaws to disrupt service[5]. At-
tackers employ hijacked devices and botnets to broaden the reach and impact of their attacks.
Mitigation strategies include implementing DDoS protection measures such as traffic filter-
ing and rate limiting, monitoring for unusual network traffic patterns, ensuring infrastructure
is scalable to handle increased traffic loads, establishing redundancy and failover mecha-
nisms, and encouraging stakeholder collaboration to share threat intelligence and coordinate
response efforts.

2.3. Firmware Exploitation

By focusing on firmware defects, IoT device firmware becomes exposed to attack. Attackers
use these holes to gain unauthorised access to devices, steal sensitive data, or install harm-
ful malware. By exploiting firmware defects, attackers can circumvent security safeguards
and risk the devices’ functionality and integrity. Regular firmware updates, fast patching
of known defects, secure boot processes, code inspections, and the usage of robust authen-
tication and access control techniques are all part of the firmware exploitation mitigation
strategy. Furthermore, to limit the risk of firmware exploitation, manufacturers should pri-
oritise robust security practises throughout the development phase.

4
2.4. Data Privacy and Breaches

Data privacy and breaches in the context of the Internet of Things refer to the protection
of private data collected and sent by IoT devices, as well as the prevention of unauthorised
access, disclosure, or use of such data. Because they routinely handle sensitive and pri-
vate data, IoT devices are enticing targets for attackers. Breaches can occur due to poor
encryption, insecure communication lines, insufficient security systems, and compromised
equipment. To deal with data privacy and breaches, strong security protocols, encryption
techniques, secure storage and transmission systems, user consent processes, and depend-
able access limitations must be implemented. To secure data privacy in IoT environments, it
is also critical to perform frequent security assessments, define data protection policies, and
implement data protection controls.

2.5. Credential Attacks

Attackers exploit weak or default usernames and passwords to gain unauthorised access to
IoT devices or connected systems. Attackers use a number of tactics, such as brute-force
attacks or credential stuffing, to repeatedly guess or use stolen credentials. Once the de-
vices have been taken over, the attackers have access to sensitive information and can even
launch more attacks. To combat credential attacks, strong authentication techniques such as
difficult passwords, two-factor authentication, and account lockouts after several failed lo-
gin attempts must be implemented. Credential attacks must be prevented by often updating
and changing default credentials, monitoring login activity, and educating users on secure
password management.

3. Intrusion Detection System

An Intrusion Detection System (IDS) is a device or software application that monitors a net-
work for malicious activity or policy violations. [6] IDSs can play a crucial role in safeguard-
ing IoT systems against cyberattacks by creating a second wall of defence, complementing
conventional preventive security measures such as authentication, authorization, and encryp-
tion. [7]

The Intrusion Detection System (IDS) serves as a defense mechanism that can promptly
inform users or block attacks upon the detection of network intrusions. In the domain of

5
cloud service providers (CSPs), IDS adoption has witnessed a rise in order to safeguard
the cloud servers used by their tenants. [8] Consequently, the network security industry
emphasizes the necessity of intrusion detection systems to prevent unauthorized access to
sensitive information. IDS actively monitors both inbound and outbound traffic and activities
to identify potential intrusions. Moreover, IDS employs machine learning (ML) algorithms
to enhance its capabilities. ML algorithms automate the process of analyzing data, enabling
precise identification of likely intrusion attempts [9]. Over the past two decades, IDS has
exhibited rapid growth and provides advanced tools and technologies for numerous fields
[10].

3.1. Importance of IDS

In 2016, a massive distributed denial-of-service (DDoS) attack took down major websites,
including Twitter, Netflix, and Reddit, and was orchestrated using the Mirai botnet [11]. The
Mirai botnet is a type of Internet of Things (IoT) botnet that is largely made up of IoT devices
such as digital cameras and DVR players. The botnet was created by infecting IoT devices
with malware and coordinating them to bombard a server with traffic until it collapsed under
the strain. The Mirai botnet was able to infect hundreds of thousands of IoT devices, making
it one of the largest botnets in history.

The Mirai botnet attack is significant because it highlights the vulnerability of IoT devices
to cyberattacks and the need for effective intrusion detection systems (IDSs) to safeguard
IoT systems against such attacks [12]. IDSs can play a crucial role in identifying security
incidents and analyzing the quantity and types of attacks [13]. However, the Mirai botnet
attack also underscores the challenges associated with IoT security, including the need to
monitor for threats by means of an IDS and the need to promptly react to malicious activities
to prevent private data from leaving the network.[14]

3.2. IDS Types

Intrusion Detection Systems (IDS) are software programs that collect and evaluate various
security metrics or parameters related to a network. IDSs keep track of network activity and
perform thorough network analyses to detect potential assaults and unwanted network access.
Having the system type in context, IDSs can be categorized into three major methodological
ideas: misuse detection, anomaly detection, or a combination of the two. Misuse detection
compares the attack behaviours used to break into systems with the user activity that has
been recorded. Anomaly detection, on the other hand, detects behaviour that is considerably

6
 Figure 3.1: OMG botnet, a variant of Mirai, can turn IoT devices into proxy servers for
malicious activities, including DDoS attacks.

different from normal activity. IDSs can detect intrusions from outside the network (outsider
assaults) or from authorized users who want more power (insider attacks). Privileged users
can also try to abuse their power.

Figure 3.2: Different types of Intrusion Detection Systems.[15]

7
Intrusion Detection Systems (IDS) can also be categorized based on their invasive behaviour,
with two primary types being network-based IDS (NIDS) and host-based IDS (HIDS). NIDS
focuses on monitoring network activity, which is achieved through network monitoring
equipment like switches, routers, and other network devices. While NIDS actively moni-
tors every packet within the network, HIDS relies on log files and user account information.
It is common for organizations to adopt a combination of both NIDS and HIDS, allowing for
comprehensive network security coverage. Based on the penetration strategy,

3.3. Current Strategies and Limitations

Machine learning and deep learning methods, such as convolutional neural networks, have
been used to improve the ability of IDSs to detect and classify intrusions in IoT networks.
Researchers have proposed various IDSs for IoT, including K-means clustering unsupervised
learning-based IDS, decision tree-based supervised IDS, and a hybrid two-stage IDS that
combines K-means and decision tree-based IDS.

Despite the significant contributions of machine learning (ML) and deep learning (DL) in
addressing real-world problems, there are various limitations that need to be considered.
These constraints include:

1. Privacy Concerns: ML and DL models typically require the user to upload their
private data to a central entity for training the centralized model. This raises privacy
concerns as sensitive information may be exposed. Users may be reluctant to share
their data, especially when it involves personal or confidential information.

2. Performance and Scalability Issues: As the scale of the network increases, the per-
formance of ML and DL models may decrease. Centralized processing introduces a
single point of failure, undermining the integrity and quality of services (QoS). The
network may experience bottlenecks and congestion, leading to delays and reduced
performance.

3. Time-Consuming Centralized Processing: Intrusion detection systems (IDS) require

fast analysis to quickly detect and respond to security threats. However, the central-
ized processing involved in ML and DL can be time-consuming, resulting in delays in
identifying and mitigating intrusions.

4. Sensitivity of IoT Data: IoT devices often collect and process data from end-users.
This data may include personal or sensitive information that needs to be protected.

8
 The centralized nature of ML and DL models can potentially expose this sensitive
data, making it vulnerable to security breaches.

5. Burdening and Costly Data Collection in 5G/6G Networks: Collecting data in

5G/6G networks can be challenging and costly. These networks consist of highly
diverse data types, such as text, audio, video, and augmented/virtual reality (AR/VR).
Gathering and processing this diverse data can be burdensome and resource-intensive.

As a result, alternative approaches are being explored to address these limitations. One
such approach is decentralized or federated learning. Federated learning allows models to
be trained locally on IoT devices, without the need to upload raw data to a centralized en-
tity. This approach provides privacy preservation while addressing the scalability and per-
formance issues associated with centralized processing. Additionally, advancements in edge
computing and distributed learning techniques are being explored to overcome the challenges
of IoT data collection and processing in 5G/6G networks. These approaches aim to enhance
the security, privacy, and efficiency of IDS in IoT systems.

4. Federated Learning

Due to the rapid progress in computational capabilities, digital devices are generating an
enormous amount of data continuously. This over-arching generation and transmission of
data consequently are coupled with the threat to the protection of personal and confiden-
tial information. So the desire for limiting the transmission of data from the end devices
to centralized servers of the users has called for the emergence of the concept of Federated
Learning(FL). FL enables the training of Machine learning (ML) models locally over remote
or edge devices. It was first conceptualized in a workshop arranged by Google’s research,
’Private Multi-Party Machine Learning’ (2016) [16]. Primarily, the techniques of the fed-
erated learning paradigm are a synergy of 3 existing research areas: distributed systems,
machine learning, and privacy. The current data-driven world has an avid need for high com-
putational capacity with an immediate and frequent need for locally accessible data. So the
centralized way of collecting, generating, and accessing data has very little capacity to cater
to the processing need for the current manifestation of technological operation. The concept
of distributed processing systems has evolved to meet the growing demand for parallel han-
dling operations and speedup of the systems[17]. But architectures and algorithms that allow
distributed processing also call for the need of comprehending the trade-off between com-
munication and efficiency. Artificial Intelligence(AI) is a branch of the computer science

9
paradigm which aims to develop systems that have processing capacity parallel to human
intelligence. Machine learning is at the core of AI which emerged to impart learning and
decision-making abilities to computing systems. The materialization of these technological
developments largely depends on data that are highly private and confidential. Hence there
is always a need for balancing the potency of AI and respecting the privacy of the data that
is being used as well. Although the primary focus of federated learning is to protect user pri-
vacy and confidentiality of data, it has certain distinctiveness in contrast to the conventional
privacy-preserving algorithms.

Figure 4.1: Federated Learning [18]

The conventional Deep Neural Network is provided with a training dataset D that consists
of input data points xi and corresponding class labels ci . The input data points xi contain
a number of features fi , and the class labels ci belong to a set of the class values in the
dataset ci . The DNN aims to trace the relationship between the input data points and their
corresponding class labels so that it can make predictions on unseen data.

The DNN architecture consists of multiple layers, with the final layer containing nodes of the
same number as classes in the set C . Each node in the final layer corresponds to a different
class in C and is used to compute the predicted probability that an input data point belongs
to that class.

The DNN is trained using a loss function L , which measures the difference between the
predicted output of the DNN and the expected output. The loss function is calculated by
summing the loss for each input data point in the training set D and averaging the result.
Finding the values of the model parameters θ that minimize the loss function is the aim of
training.

After the training, the DNN has the ability to predict unseen data by computing the predicted
probabilities for all the nodes in the last layer which represents the classes of the dataset and

10
predicts the class with the largest probabilistic value as the correct class.

A centralized machine learning system is the traditional approach for optimizing models,
where the entire data is collected and stored in a central server, and the model training is
performed on this data.

In a centralized machine learning system, the training data is typically collected and stored
in a single database or server. The model is then optimized on this centralized data using
a batch learning process, where the entire dataset is used to train the model at once. This
can be done using gradient descent, ADAM, or any other optimization technique. Once the
model is optimized, it is deployed for making predictions on new instances. One of the main
disadvantages of a centralized machine learning system is that it allows for easy accessibility
to a significant amount of data. In addition to that, the need to accumulate and reserve
large amounts of data in a single location can be costly and may raise privacy concerns.
Additionally, the model can have difficulty in generalization, as it has only been exposed to
the centralized dataset.

In the FL approach, the data is collected from a distributed set of devices or machines that
can be used to build models, without sending any data to a central location. This is useful
in situations where the data is sensitive or private, and cannot be transmitted to a central
location for processing.

In the current federated learning frameworks, each client device possesses its own specific
dataset. These client devices utilize their local datasets to train a model independently, and
subsequently contribute their updates towards the development of a global model. To do this,
the devices or machines perform the training process in their individual devices and transmit
the local model to an aggregator. The aggregator combines the updated parameters from all
of the devices or machines to calculate a global model, which is used to make predictions.

4.1. Federated Learning Steps

A FL framework generally has 3 key steps [19] in each round:

11
 Figure 4.2: Steps of Federated Learning [19]

• Initialization and Client Selection: Before starting the rounds, the central aggrega-
tor selects some of the participants and initializes the system. Several methods and
algorithms are proposed for the optimum initialization and client selection process.
The process of selecting clients in various algorithms takes into account factors such
as network conditions, the quality of local updates, and the training frequency of the
clients. Numerous proposed algorithms illustrate how these criteria are utilized in the
client selection process.

• Local Training: Once the initialization phase is complete, the chosen clients under-
take training within their respective local environments using the model provided by
the server. They then calculate an update by minimizing the loss within their own
dataset.

• Model Aggregation, Parameter Broadcast and Download: After local training, the
model parameters attained are transmitted to the aggregator. The aggregator accumu-
lates the local updates transmitted by the selected participants and computes an up-
dated global model accumulating the knowledge learned by the different clients from
their training.The most popular mode aggregation method is the Fed-Avg [20]:

1
ω t+1 = ∑ ωkt+1
C kεCt

12
 where ω is the global model weights, C is the total number of clients Ct is the subset
of the clients participating in that particular round. Besides FedAvg, FedProx [21],
Robust Aggregation for Federated Learning [22], FedBuff [23], FedAvgM [24] etc
are some of the other aggregation methods proposed in the recent years. If the global
model converges, the model is saved and broadcast to all the clients in the network.
Upon the receival of the parameters, the local clients download them and use them to
optimize the local models for the next round.

5. Federated Learning for Intrusion Detection System

The application of Federated Learning in Internet of Things scenarios has gained significant
interest in recent years due to its benefits compared to traditional centralized learning ap-
proaches. By leveraging the power of Federated Learning, IDS in IoT networks can benefit
from the collective knowledge and data of multiple devices without compromising data pri-
vacy. This distributed approach enables the creation of more effective and scalable intrusion
detection systems, particularly in IoT networks with limited resources and restricted con-
nectivity. FL for IDS offers the potential to improve security and privacy in IoT networks
by enhancing anomaly detection and enabling collaborative and decentralized intrusion de-
tection. The implementation of federated learning in IDS involves two main phases: local
update and global aggregation. In the local update phase, each device or node in the network
trains its own local model using its local data. This ensures that data privacy is maintained,
as the data remains on individual devices and is not shared. In the global aggregation phase,
the locally trained models are aggregated or combined to create a global model that rep-
resents the knowledge learned from all the devices in the network. This global model can
then be used for intrusion detection and identifying unwanted intrusions in the network. By
leveraging the power of federated learning, IDS can benefit from the collective knowledge
and data of multiple devices without compromising data privacy. This distributed approach
enables the creation of more effective and scalable intrusion detection systems, particularly
in IoT networks where the devices may have limited resources and restricted connectivity.

5.1. FL for Heterogeneous Anomaly Detection

Federated Learning (FL) offers a solution to the challenges faced by anomaly-based Intrusion
Detection Systems (IDS). In realistic scenarios, IDS often generate false alarms due to insuf-
ficient and unlabelled data, as well as inefficient algorithms. Attackers can modify intrusion

13
attempts to evade IDS through techniques like traffic masking and network steganography.

To overcome these limitations, FL allows access to data available on client devices, providing
real-time protection in larger networks. [25] proposed a disagreement-based semi-supervised
learning approach using FL to tackle data size limitations and handle unlabelled data, leading
to better prediction accuracy and fewer false alarms compared to supervised ML algorithms.

In multinational companies, managing the protection of each network node is challenging.

Different teams and departments have varying data, communication, and bandwidth require-
ments, resulting in significant traffic flow variations. The heterogeneity of data poses diffi-
culties for a centralized model, and the complexity of a federated architecture can address
this issue.

Experiments by [26] assessed the usefulness of heterogeneous and non-IID data in FL, high-
lighting its superiority over centralized architectures while acknowledging challenges posed
by the non-IID nature of the data.

FL offers promising solutions for IDS by addressing data limitations, heterogeneity, and pri-
vacy concerns. It enables collaborative learning while providing better prediction accuracy
and reducing false alarms. Further research is needed to overcome challenges associated
with non-IID data and optimize FL algorithms for enhanced IDS performance.

5.2. FL for DDoS Attacks

The use of Federated Learning (FL) for Distributed Denial of Service (DDoS) attack detec-
tion is gaining attention as a means to ensure the Quality of Service (QoS) for end-users.
DDoS attacks pose a significant threat to the Internet by overwhelming and exhausting target
systems through coordinated attacks from various computer systems.

Although FL for Intrusion Detection Systems (IDS) is still in its early stages, researchers
are increasingly exploring its potential. [27] proposed an FL-based mitigation architecture
called FLEAM in the Industrial Internet of Things (IIoT) context. FLEAM utilizes FL, fog
computing, and cloud computing to effectively handle large volumes of DDoS attacks. In
this approach, each fog node locally downloads a training model, retrains it, and uploads the
model parameter to the cloud.

Experimental results indicate that FLEAM significantly reduces the mitigation delay by ap-
proximately 72% and improves accuracy by 47% compared to traditional solutions. This
highlights the potential of FL in enhancing DDoS attack detection and mitigating their im-

14
pact on system performance and end-user experience.

Overall, FL shows promise in improving the effectiveness and efficiency of DDoS attack
detection and mitigation, particularly in the IIoT context. Further research and development
are required to optimize FL-based approaches and ensure their practical implementation in
real-world scenarios.

5.3. FL for privacy-preserving data handling

In the context of privacy-preserving data handling, Federated Learning (FL) offers a solution
for overcoming limitations faced by traditional Deep Anomaly Detection (DAD) systems
used in Intrusion Detection Systems (IDS). The main goal of an IDS is to ensure the con-
fidentiality, integrity, and availability of information against external intrusions. However,
training an anomaly detection statistical model requires a significant amount of data, and the
collection of such data can compromise privacy. Privacy preservation becomes challenging
when traditional DAD systems store a large amount of unprotected data in a central location,
such as a server, making it vulnerable to breaches.

FL addresses these privacy concerns by allowing data to remain on local devices and avoiding
the need to transmit sensitive information to a central server. This decentralized approach
enables model training while preserving the confidentiality of data, making FL a suitable
framework for privacy-preserving data handling in IDS.

FL integration ensures privacy while training anomaly detection models in IDS. By avoiding
the centralization of data and the use of vulnerable communication channels, FL mitigates
the privacy risks associated with traditional DAD systems. Further research and development
are necessary to optimize FL-based approaches and ensure their practical implementation in
real-world scenarios while upholding data confidentiality.

5.4. FL for IDS in low-power IoT devices

Federated Learning (FL) offers a solution for Intrusion Detection Systems (IDS) in low-
power Internet of Things (IoT) devices. Low-power IoT devices, including edge devices, are
prime targets for data-hungry attackers due to the confidential and sensitive data they collect.
These devices have limited processing capabilities and are often designed with low-security
measures due to their constrained power allocation.

FL utilizes the distributed architecture of the network to leverage the overall computation ca-

15
pability, creating a stronger and robust IDS against intruders. By intercepting and carefully
processing network packets, anomalies in device behavior can be identified. [28] introduced
DÏoT, a federated self-learning anomaly detection system specifically designed for IoT de-
vices. It utilizes device-type-specific communication profiles without human interaction and
a pre-labelled dataset for model training.

In the DÏoT system, data packets are represented as symbols in a predefined language, allow-
ing language analysis techniques to effectively differentiate anomalous packets from normal
ones. Each communication node in the network houses an anomaly detection system, and
model aggregation occurs on a predefined server. The packets entering each device are cap-
tured, and their characteristic features are extracted and mapped to symbols. These symbols
undergo probability computation using a recurring unit, such as GRU, and the packets are
classified as anomalous or normal.

DÏoT has been tested against the Mirai attack and demonstrated its self-learning capability,
adapting to device behavior over time. The system achieves 95.6% accuracy in detecting
anomalies with an average processing time of 257 ms, making it suitable for low-power IoT
devices and sensor networks requiring fast IDS capabilities. Additionally, DÏoT provides
low false alarm rates in real-time scenarios.

5.5. Challenges in Deploying FL in IDS

Although federated learning (FL) has showcased its superiority over conventional deep learn-
ing (DL) models in terms of intrusion detection system (IDS) deployment, it is not without
its drawbacks and vulnerabilities. The key challenges and issues of using FL for IDS are
discussed below.

1. Communication Overhead: The cost of communication per training round is a pri-

mary limitation of FL. Transmitting model parameters and trained models between
the central server and clients incurs traffic load, packet loss, and variable communi-
cation time. This limits the overall throughput of a federated network, particularly in
intrusion detection scenarios that require efficiency and speed.

2. Federated Poisoning Attacks: The distributed nature of data in client devices, while
protecting data privacy, is vulnerable to poisoning attacks. Clients can modify data la-
bels, leading to poisoned predictions and a compromised global model. These attacks
have been demonstrated using Generative Adversarial Networks (GAN) and have the
potential to cause significant damage.

16
 3. High False Alarms with non-IID data: FL relies on large amounts of data, but the
heterogeneity and non-IID nature of the data can result in improperly trained local and
global models. This leads to a high number of false alarms in intrusion detection due
to the variability in data attributes and the complexity of the classification task.

4. Resource Management in Low Power IoT Devices: Low-power IoT edge devices lack
computation power for continuous DL model training. Training locally on these de-
vices may drain their energy quickly and introduce latency to the server. Energy-
efficient algorithms are needed to implement FL in low-power devices.

5. Vulnerabilities in IDS Setups: IDS deployed using FL and DL algorithms are suscep-
tible to various vulnerabilities that can compromise the confidentiality and integrity of
the system.

6. Precise Deployment of IDS: DL models have numerous parameters and hyper-parameters

that are difficult to fine-tune in a federated learning scenario. The heterogeneity of
client data requires customized parameters for optimal performance, which is chal-
lenging to achieve in IDS applications with unpredictable attack types and sources.

7. Precise Deployment of IDS: DL models have numerous parameters and hyper-parameters

that are difficult to fine-tune in a federated learning scenario. The heterogeneity of
client data requires customized parameters for optimal performance, which is chal-
lenging to achieve in IDS applications with unpredictable attack types and sources.

In summary, FL for IDS has limitations related to communication overhead, poisoning at-
tacks, false alarms, resource management in low-power devices, vulnerabilities in IDS se-
tups, and the precise deployment of IDS models.

17
6. A Case Study: Federated IoT Attack Detection Using Decentralized
Edge Data [1]

A proposed method by Regan et al. [1]incorporates a decentralized approach, leveraging a

deep autoencoder to detect botnet attacks by utilizing on-device traffic data. The proposed
federated technique addresses the privacy issue by restricting data transfers or movements
off the network edge. Instead, the machine learning algorithm is sent straight to the data
source that includes extra data security as well.

6.1. Proposed Approach

6.1.1. Architecture

For botnet attack detection, the suggested architecture is a decentralised federated learning
(FL) solution with a deep autoencoder anomaly detection engine. IoT security gateways use
port mirroring and packet sniffing to run and monitor network traffic. These gateways house
the FL and anomaly detection processes. They exchange data with a FL server in order to

Figure 6.1: Overview of proposed architecture.

choose and obtain deep autoencoder models for specific IoT devices. The FL server transmits
global models to the virtual workers of the gateways, who train local models. The gateways

18
return trained models for aggregation, gradually increasing the FL quality. The overall ar-
chitecture is shown in Figure 6.1, that is compatible with a wide range of hardware, devices,
and settings. Each security gateway can host a number of virtual employees, resulting in a
many-to-one relationship.

6.1.2. Deep autoencoder design

The deep autoencoder used in this investigation, which has shown excellent accuracy in IoT
intrusion detection systems [29]. Through numerous encoding layers and an Artificial Neural
Network (ANN), it compresses input data representing network traffic to an encoded state.
The input dimensions are reduced during encoding, and the data is returned to its original
dimensions during decoding. When regenerated data exceeds a specific threshold, malicious
activities can be identified.

Figure 6.2: The autoencoder implementation.

In Figure 6.2,the implementation of the deep autoencoder is shown that consists of an input
layer, a number of encoding levels, a number of decoding layers, and an output layer. While
the decoding layers reverse the compression, each encoding layer gradually reduces the data
size. Input is recreated and compared to the original in the output layer.

19
6.1.3. Security gateways

The proposed FL-based solution comprises of a software service that operates on specialised
security gateways located at the network periphery before IoT devices. Gateways monitor
UDP and TCP traffic without interfering with network activity. Network port mirroring is
used to replicate certain traffic data for anomaly detection. A separate security service gets
warnings of discovered anomalies for management or human action. The design assumes
that the computational and software limitations of IoT device hardware and software pre-
clude them from assisting the machine learning process.

6.1.4. Multi-virtual workers process

In 6.3, virtual worker architecture in federated learning has been shown. The virtual worker,
which represents an edge device, remains in the simulation region for the length of the ex-
periment. It is provided a model to work with, and it returns the model upon request. A
Tensor stores the virtual worker’s position and data. It employs a linear layer to construct
a deep autoencoder model from a batch of pre-processed data. Following training, the vir-
tual worker sends its model to the aggregation service for processing. The model aggregator
awaits the delivery of all virtual workers’ models before aggregation. After extracting the
loss and zeroing out the gradient optimizer, the next training cycle begins. These are the
main duties that virtual employees must do throughout their training.

6.1.5. Federated learning simulation and implementation

The PySyft framework for FL and PyTorch are used in the suggested implementation. To
find abnormalities, the deep autoencoder model is trained on normal traffic data. The ability
to reconstruct regular data and the inability to do so point to abnormalities in the presence
of malicious traffic. Profile requests, the construction of virtual workers, the distribution of
batched data, epoch-based training, gradient optimisation, the aggregation of local models,
and the gathering of output loss from the global model are all part of the process. The pro-
gramme keeps running forever while improving accuracy. A threshold is created for testing
when all of the epochs have been completed. The method uses distributed virtual workers to

20
 Figure 6.3: Virtual worker architecture in federated learning environment.

train the anomaly detection model in a safe and private manner.

6.2. Research methodology

6.2.1. Dataset & Performance metrics

This paper made use of an IoT botnet dataset[30] that is often utilised in academic inves-
tigations. It is composed of diverse collections of both good and harmful traffic collected
from infected devices. Each device’s network traffic was inspected, and statistics and prop-
erties such as source and destination IP addresses, MAC-IP relationships, and TCP/UDP
connections were recorded. The preprocessing approach involved feature extraction, dataset
separation into training and testing sets, and scaling with a common scalar. Using the training

21
data, an autoencoder for a neural network with particular hyperparameters was developed.
Malicious data was chosen for testing and combined with the benign testing set to create a
mixed testing dataset for evaluation.

The assessment metrics for FL-aided anomaly detection include true positive (TP), false
positive (FP), true negative (TN), false negative (FN), threshold (TR), mean squared error
(MSE), and standard deviation (STD). The training metrics include threshold calculation,
false positive rate calculation, and loss function reduction. During the testing phase, a con-
fusion matrix is used to determine accuracy, precision, recall, and F-measure. PyTorch is
used to build the model, the mean squared error criterion is employed, stochastic gradient
descent is used as the optimizer, and the false positive rate is calculated. During testing, the
threshold is critical for distinguishing between legal and malicious transmissions.

6.2.2. Evaluation and testing goals

After defining a threshold, a virtual worker in evaluation mode evaluates the model’s quality.
The mean squared error is determined between the evaluation data and the anticipated data,
aggregating numbers over the threshold to count false positives. The assessment set size is
compared to the number of false positives. During the testing phase, accuracy, precision,
recall, and F-measure scores are expected to be obtained. The baseline strategy is tested as
a non-FL approach using the same deep autoencoder technology. The FL plan calls for the
addition of numerous virtual staff. Both training and testing experiments employ the same
dataset,batch sizes, learning rate, epoch count, and hardware to ensure consistency.

6.3. Results

The evaluation’s conclusions about how well the suggested federated anomaly detection sys-
tem performs are intriguing. The federated method regularly delivers fewer false positive
rates, with a considerable improvement seen, when compared to the non-FL baseline(Figure
6.4).

After training with 55 features, the accuracy of the federated solution exceeds expectations,
reaching 99%. Precision is consistently between 0.97 and 0.98, suggesting accurate identi-

22
 (a) Detected False Positives (b) Average Accuracy

Figure 6.4: Fedarated vs Non-fedarated baseline

fication of malicious assaults. The federated strategy outperforms the baseline in terms of
both the F-measure and recall scores, which both show a constant rising trend. The federated
solution achieves a remarkable accuracy of 94.04% on a typical basis, outperforming the
non-FL baseline’s accuracy of 93.73%. The ability to train on decentralised data and dis-
tributed computations is made possible by federated learning, which improves efficiency in
identifying various network abnormalities and botnet assaults. These outcomes demonstrate
the federated solution’s efficacy in various traffic conditions and its scalability and practical
application possibilities.

7. Conclusion

Federated Learning offers a decentralized and privacy-preserving approach to machine learn-

ing, making it well-suited for Intrusion Detection Systems in IoT networks. By training mod-
els on decentralized data across multiple edge devices, FL enables collaborative and scalable
intrusion detection without compromising user privacy. Further research and experimenta-
tion are being conducted to optimize methodologies for implementing Federated Learning
in IDS and improve the security and privacy of IoT networks.

23
8. References

[1] Christopher Regan, Mohammad Nasajpour, Reza M Parizi, Seyedamin Pouriyeh, Ali
Dehghantanha, and Kim-Kwang Raymond Choo. Federated iot attack detection using
decentralized edge data. Machine Learning with Applications, 8:100263, 2022.

[2] Enrique Mármol Campos, Pablo Fernández Saura, Aurora González-Vidal, José L
Hernández-Ramos, Jorge Bernal Bernabé, Gianmarco Baldini, and Antonio Skarmeta.
Evaluating federated learning for intrusion detection in internet of things: Review and
challenges. Computer Networks, 203:108661, 2022.

[3] Elisa Bertino and Nayeem Islam. Botnets and internet of things security. Computer,
50(2):76–79, 2017.

[4] MohammadNoor Injadat, Abdallah Moubayed, and Abdallah Shami. Detecting botnet
attacks in iot environments: An optimized machine learning approach. In 2020 32nd
International Conference on Microelectronics (ICM), pages 1–4. IEEE, 2020.

[5] Ruchi Vishwakarma and Ankit Kumar Jain. A survey of ddos attacking techniques and
defence mechanisms in the iot network. Telecommunication systems, 73(1):3–25, 2020.

[6] Hung-Jen Liao, Chun-Hung Richard Lin, Ying-Chih Lin, and Kuang-Yuan Tung. In-
trusion detection system: A comprehensive review. Journal of Network and Computer
Applications, 36(1):16–24, 2013.

[7] Ansam Khraisat, Iqbal Gondal, Peter Vamplew, and Joarder Kamruzzaman. Survey
of intrusion detection systems: techniques, datasets and challenges. Cybersecurity,
2(1):1–22, 2019.

[8] Salman Iqbal, Miss Laiha Mat Kiah, Babak Dhaghighi, Muzammil Hussain, Suleman
Khan, Muhammad Khurram Khan, and Kim-Kwang Raymond Choo. On cloud security
attacks: A taxonomy and intrusion detection and prevention as a service. Journal of
Network and Computer Applications, 74:98–120, 2016.

[9] P Amala, G Gayathri, S Dinesh, and S Prabagar. Effective intrusion detection system
using support vector machine learning. Int. J. Adv. Sci. Eng. Res, 3:302–305, 2018.

[10] Lukman Hakim, Rahilla Fatma, et al. Influence analysis of feature selection to network
intrusion detection system performance using nsl-kdd dataset. In 2019 International
conference on computer science, information technology, and electrical engineering
(ICOMITEE), pages 217–220. IEEE, 2019.

24
[11] Jasper Manuel, Rommel Joven, and Dario Durando. Omg: Mirai-based bot turns iot
devices into proxy servers. Fortinet. doi, 10(3395352.3402619), 2018.

[12] Majda Wazzan, Daniyal Algazzawi, Omaima Bamasaq, Aiiad Albeshri, and Li Cheng.
Internet of things botnet detection approaches: Analysis and recommendations for fu-
ture research. Applied Sciences, 11(12):5713, 2021.

[13] Anamika Chauhan, Rajyavardhan Singh, and Pratyush Jain. A literature review: Intru-
sion detection systems in internet of things. In Journal of Physics: Conference Series,
volume 1518, page 012040. IOP Publishing, 2020.

[14] Rhishabh Hattarki, Shruti Houji, and Manisha Dhage. Real time intrusion detection
system for iot networks. In 2021 6th International conference for convergence in tech-
nology (I2CT), pages 1–5. IEEE, 2021.

[15] Hanan Zainel and Cemal Koçak. Lan intrusion detection using convolutional neural
networks. Applied sciences, 12(13):6645, 2022.

[16] Jakub Konečný, H. Brendan McMahan, Felix X. Yu, Peter Richtarik, Ananda Theertha
Suresh, and Dave Bacon. Federated learning: Strategies for improving communication
efficiency. In NIPS Workshop on Private Multi-Party Machine Learning, 2016.

[17] Leonard Kleinrock. Distributed systems. Communications of the ACM, 28(11):1200–

1213, 1985.

[18] Muhammad Habib ur Rehman, Khaled Salah, Ernesto Damiani, and Davor Svetinovic.
Towards blockchain-based reputation-aware federated learning. In IEEE INFOCOM
2020-IEEE Conference on Computer Communications Workshops (INFOCOM WK-
SHPS), pages 183–188. IEEE, 2020.

[19] Dinh C Nguyen, Ming Ding, Pubudu N Pathirana, Aruna Seneviratne, Jun Li, and
H Vincent Poor. Federated learning for internet of things: A comprehensive survey.
IEEE Communications Surveys & Tutorials, 2021.

[20] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera
y Arcas. Communication-efficient learning of deep networks from decentralized data.
In Artificial intelligence and statistics, pages 1273–1282. PMLR, 2017.

[21] Anit Kumar Sahu, Tian Li, Maziar Sanjabi, Manzil Zaheer, Ameet Talwalkar, and Vir-
ginia Smith. On the convergence of federated optimization in heterogeneous networks.
arXiv preprint arXiv:1812.06127, 3:3, 2018.

[22] Krishna Pillutla, Sham M Kakade, and Zaid Harchaoui. Robust aggregation for feder-
ated learning. arXiv preprint arXiv:1912.13445, 2019.

25
[23] John Nguyen, Kshitiz Malik, Hongyuan Zhan, Ashkan Yousefpour, Mike Rabbat, Mani
Malek, and Dzmitry Huba. Federated learning with buffered asynchronous aggregation.
In International Conference on Artificial Intelligence and Statistics, pages 3581–3607.
PMLR, 2022.

[24] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of
non-identical data distribution for federated visual classification. arXiv preprint
arXiv:1909.06335, 2019.

[25] Wenjuan Li, Weizhi Meng, and Man Ho Au. Enhancing collaborative intrusion detec-
tion via disagreement-based semi-supervised learning in iot environments. Journal of
Network and Computer Applications, 161:102631, 2020.

[26] Christopher Briggs, Zhong Fan, and Peter Andras. Federated learning with hierarchical
clustering of local updates to improve training on non-iid data. In 2020 International
Joint Conference on Neural Networks (IJCNN), pages 1–9. IEEE, 2020.

[27] Jianhua Li, Lingjuan Lyu, Ximeng Liu, Xuyun Zhang, and Xixiang Lyu. Fleam: A
federated learning empowered architecture to mitigate ddos in industrial iot. IEEE
Transactions on Industrial Informatics, 18(6):4059–4068, 2021.

[28] Thien Duc Nguyen, Samuel Marchal, Markus Miettinen, Hossein Fereidooni,
N Asokan, and Ahmad-Reza Sadeghi. Dı̈ot: A federated self-learning anomaly de-
tection system for iot. In 2019 IEEE 39th International conference on distributed com-
puting systems (ICDCS), pages 756–767. IEEE, 2019.

[29] Yisroel Mirsky, Tomer Doitshman, Yuval Elovici, and Asaf Shabtai. Kitsune: an
ensemble of autoencoders for online network intrusion detection. arXiv preprint
arXiv:1802.09089, 2018.

[30] Yair Meidan, Michael Bohadana, Yael Mathov, Yisroel Mirsky, Asaf Shabtai, Dominik
Breitenbacher, and Yuval Elovici. N-baiot—network-based detection of iot botnet at-
tacks using deep autoencoders. IEEE Pervasive Computing, 17(3):12–22, 2018.

26