HOMEFUS: A Privacy and Security-Aware Model for IoT Data Fusion

in Smart Connected Homes

Kayode S. Adewole

1,2 a

and Andreas Jacobsson

1,2 b

Department of Computer Science and Media Technology, Malm

o University, Malm

o, Sweden

Internet of Things and People Research Center, Malm

o University, Malm

o, Sweden

Keywords:

Smart Homes, Internet of Things, Data Fusion, Security, Privacy, Federated Learning, Sensors Selection.

Abstract:

The beneﬁt associated with the deployment of Internet of Things (IoT) technology is increasing daily. IoT

has revolutionized our ways of life, especially when we consider its applications in smart connected homes.

Smart devices at home enable the collection of data from multiple sensors for a range of applications and

services. Nevertheless, the security and privacy issues associated with aggregating multiple sensors’ data

in smart connected homes have not yet been sufﬁciently prioritized. Along this development, this paper

proposes HOMEFUS, a privacy and security-aware model that leverages information theoretic correlation

analysis and gradient boosting to fuse multiple sensors’ data at the edge nodes of smart connected homes.

HOMEFUS employs federated learning, edge and cloud computing to reduce privacy leakage of sensitive

data. To demonstrate its applicability, we show that the proposed model meets the requirements for efﬁcient

data fusion pipelines. The model guides practitioners and researchers on how to setup secure smart connected

homes that comply with privacy laws, regulations, and standards.

1 INTRODUCTION

Internet of Things (IoT) has paved ways for connect-

ing different sensors and smart devices to beneﬁt from

a range of applications and services including smart

health, smart grids, intelligent transportation, smart

manufacturing, autonomous driving, and smart agri-

culture. With over 7 billion connected IoT devices to-

day, the number of smart devices that will be powered

by IoT technology is expected to grow to 22 billion

by 2025 (Gartner Inc., 2017). The vision of IoT is to

seamlessly connect everything in the physical world

over the Internet. This technology has transcended

into our homes, allowing us to connect smart home

devices using sensors, actuators, and controllers that

are equipped with wireless connectivity, and cognitive

computing technologies (Rahman et al., 2016). The

main challenge is how to ensure privacy and security

while also complying with laws, regulations and stan-

dards.

There are privacy laws and regulations such as the

General Data Protection Regulation (GDPR) in Eu-

rope, Health Insurance Portability and Accountability

Act (HIPAA) in the United States, Japan Personal In-

https://orcid.org/0000-0002-0155-7949

https://orcid.org/0000-0002-8512-2976

formation Protection Act, and China’s Cybersecurity

Law. For instance, Article 23 of the GDPR enforces

data controllers to store and analyze only the neces-

sary data required to achieve data collection objec-

tives, and to limit access to sensitive data to autho-

rized entities (A

ıvodji et al., 2019). GDPR mandates

data controllers to obtain informed consent from data

subjects before collecting and analyzing their data,

and to provide them with the right to access, review,

correct, and delete their data. Despite the existence of

laws, regulations and standards, IoT companies con-

tinue to release devices into the market some of which

are vulnerable (Bugeja et al., 2019; Rahman et al.,

2016). Many users of IoT devices lack the knowledge

of the amount and the type of data collected by the

devices, and to what extent they are being utilized.

This makes it difﬁcult for users to make informed de-

cisions regarding their privacy and how to control the

data collection and distribution process.

The privacy and security risks are further com-

plicated when multiple sensors are collaborating and

when the data collected have to be transmitted to the

cloud via the Internet (Mena et al., 2022; Chimamiwa

et al., 2021; Adewole and Torra, 2022b; Adewole

and Torra, 2022a). For instance, activities recogni-

tion in smart homes may involve data collection about

Adewole, K. and Jacobsson, A.

HOMEFUS: A Privacy and Security-Aware Model for IoT Data Fusion in Smar t Connected Homes.

DOI: 10.5220/0012437900003705

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 9th International Conference on Internet of Things, Big Data and Security (IoTBDS 2024), pages 133-140

ISBN: 978-989-758-699-6; ISSN: 2184-4976

133

home occupant’s behaviors, some of which are per-

sonally identiﬁable and sensitive. Some smart home

devices, e.g., cameras, wearables, and health moni-

toring devices may not only track occupancy or ﬁt-

ness activities, but also infer sensitive information of

the users. The complexity and lengthy nature of the

policy documents that usually accompany IoT devices

from the manufacturers also make it difﬁcult for the

users to have a broad understanding of the nature of

the data the IoT devices are collecting and how these

data are being shared. With this development, align-

ing with the GDPR regulations on informed consent

is becoming very difﬁcult to implement in smart con-

nected homes (Pathmabandu et al., 2023). This re-

quires users to make adequate decision concerning

the risk associated with sharing their personal data

with the third party. Therefore, there is a need for

models that enforces privacy and security for smart

connected homes to minimize information disclosure

about users. This is the goal of HOMEFUS. More

speciﬁcally, this paper contributes in the following

ways:

1. We propose a privacy and security-aware model

that aggregates sensors’ data for effective IoT ap-

plication and service delivery in smart connected

homes.

2. We enforce privacy and security for sensors data

collection, processing, and analysis both at the

edge and in the cloud using federated learning.

3. The model leverages information theoretic corre-

lation analysis and gradient boosting for sensors

selection, fusion and modeling, which provide ef-

ﬁcient method at the edge nodes.

4. We show that the proposed model meets the re-

quirements for effective data fusion pipeline.

The remainder of this paper is organized as fol-

lows. Section 2 presents IoT ecosystems and charac-

teristics. In Section 3, related work in IoT data fusion

for smart connected homes is presented. Section 4

presents the requirements for data fusion architecture.

In Section 5, we present the proposed model. Section

6 focuses on the evaluation and risk assessment of the

model, and ﬁnally, Section 7 concludes the paper and

discusses future research directions.

2 IoT ECOSYSTEM AND

CHARACTERISTICS

The IoT is typically described as a network of het-

erogeneous devices and services. These devices are

characterised as resource-constrained, making im-

plementation of the traditional encryption and data

anonymization algorithms almost impossible. The

components of an IoT ecosystem have the capability

to communicate autonomously over the network with

unique identity. A number of architectures have been

proposed for the IoT, often with the consideration of

three layers: perception, network and application lay-

ers (A

ıvodji et al., 2019).

In Fig. 1, we present a view of an IoT ecosystem

that highlights the different components with a focus

on privacy and security. In this representation, we

add two layers thereby recognizing the crucial roles of

the users, their applications, and the controlled man-

agement of their corresponding information. Layer 1

considers ”things” in the IoT and it is the layer that

accommodates the smart connected devices. These

devices are hardware components capable of sensing,

actuating, controlling, communicating, and monitor-

ing. Layer 2 represents the communication and net-

work layer where the communication infrastructure

and standard protocols are used to establish interop-

erability among the connected smart devices of layer

1. Typical communication technologies include Zig-

bee, WiFi, Bluetooth, Cellular (e.g 5G), LoRaWAN,

and MQTT. This layer is responsible for transmitting

data to the computing and storage facilities, which are

handled by layer 3. Cloud infrastructures are mostly

leveraged in this layer. Layer 4 represents applica-

tions, which accommodates different software appli-

cations that render services to the end-users. Ser-

vices include activities recognition of home occu-

pants, occupancy sensing, energy consumption mon-

itoring, surveillance, logistics, sport, business, trans-

port, and so on.

We identiﬁed the roles of users in IoT ecosystem

which permeates across the different layers. Three

categories of users can be distinguished in connec-

tion to smart connected homes: residents, guests,

and malicious actors. The residents and guests are

considered as legitimate users. In smart connected

homes, attackers or malicious actors represent unin-

vited guests or users who aim to establish unautho-

rized connection to the home network to infer sensi-

tive information. An attacker’s goal is typically to es-

tablish unauthorized access within the IoT ecosystem

and to launch different attacks or privacy invasions.

There are also external attackers who may want to

launch complicated attacks such as denial-of-service

(DoS) and malware infections to disrupt the services

or infer some information from the target network

or smart devices on to the Internet. In layer 5, we

emphasise the need for security and privacy assess-

ment, which should consider all the layers of the IoT

IoTBDS 2024 - 9th International Conference on Internet of Things, Big Data and Security

134

ecosystem. Enforcing privacy and security in this re-

spect will help security analysts and IoT practitioners

to conduct thorough risk analysis of the IoT.

Figure 1: IoT ecosystem layers.

3 RELATED WORK

Information fusion deals with the study of explor-

ing efﬁcient automatic or semi-automatic methods of

transforming data from multiple sources to produce

improved representation for better decision making

(Ding et al., 2019). In the domain of IoT, data fusion

involves aggregation of data from multiple sensors to

extract useful information for service improvement.

This research domain has been widely studied in the

recent years. This section presents related studies in

the area of IoT data fusion with a speciﬁc focus on

smart connected home research. Data fusion in smart

connected homes can be categorized as intrusive or

non-intrusive approaches based on the types of sen-

sors aggregated. We provide the detail as follows.

3.1 Intrusive Approaches

Intrusive approaches use smart devices such as cam-

eras, wearables, GPS and microphones in smart-

phones, just to mention a few. This approach has bet-

ter accuracy for smart connected home applications,

particularly for occupant monitoring. For instance,

(Monti et al., 2022) investigate the use of multiple

cameras for occupant counting. (Kommey, 2022) de-

sign an automated ceiling fan regulator for smart con-

nected home based on the fusion of data from web

cameras and temperature sensors. (Chaaraoui et al.,

2014) proposed a weighted feature fusion approach

which relies on the use of camera data. Fusion of

passive infrared sensor that detect human presence to-

gether with contact, camera and microphones sensors

have been studied in (Chahuara et al., 2013).

Although intrusive approaches has the potential to

offer better performance for smart connected home

applications, they compromise the privacy of the

home occupants. Our proposed model addresses the

privacy challenges in the existing studies by consid-

ering federated learning and hybrid data fusion ap-

proach to reduce privacy leakage of sensitive users

information and to improve the sensor fusion method.

3.2 Non-Intrusive Approaches

The development of smart home applications that

can offer comparable performance with intrusive ap-

proaches poses research challenges. To address the

privacy concerns in intrusive monitoring, the non-

intrusive approaches consider the fusion of environ-

mental sensors, such as carbon dioxide (CO2), to-

tal volatile organic compounds (TVOC), air temper-

ature, air humidity, and smart meters. (Sayed et al.,

2023) proposed the fusion of temperature, humidity,

pressure, light level, motion, sound, and CO2 sensors

for occupancy detection in smart connected homes.

(Dutta and Roy, 2022) combined different environ-

mental sensors with contextual information to im-

prove home occupancy detection. For a detailed dis-

cussion on data fusion as well as the methods used

during the fusion pipeline, the reader is refereed to

the comprehensive review in (Ding et al., 2019).

Non-intrusive approaches reduces privacy viola-

tion of home occupant information, nevertheless, they

suffer from shortcomings in terms of efﬁciency and

reliability. Moreover, studies have shown the behav-

ior of home occupants can still be monitored through

aggregating multiple sensors data even in a non-

intrusive way (Pathmabandu et al., 2023). The ques-

tion of which sensors should be aggregated to provide

better performance while considering the limited ca-

pability of smart home devices still remain an open re-

search issues for smart connected home applications.

Additionally, risk assessment of fusion methods has

not been considered in the existing studies. Our pro-

posed model addresses the challenges in non-intrusive

domain by considering hybrid lightweight sensor se-

lection approach to improve efﬁciency and reliability

of the fusion pipeline. Particularly, HOMEFUS lever-

ages federated learning to further preserve privacy of

sensitive data in smart connected homes.

HOMEFUS: A Privacy and Security-Aware Model for IoT Data Fusion in Smart Connected Homes

135

4 REQUIREMENTS FOR SMART

HOME DATA FUSION

To beneﬁt from IoT applications in smart connected

homes and to make accurate predictions and deci-

sions, considering multiple sensors is usually advis-

able. Given the heterogeneous nature of sensors data

from multiple sources, distributing all the data over

the network will increase network bandwidth, in-

crease power consumption, decrease the longevity of

the battery-driven devices, and thus increase the costs

of the system. Therefore, in data fusion, it becomes

necessary to extract vital information from diverse ar-

rays of collected sensor data to improve data qual-

ity and facilitate decision making. The question of

which sensors should be aggregated to improve pre-

dictive performance still remain unsolved. Along this

development, (Ding et al., 2019) provides the basic

requirements for data fusion models in IoT scenarios.

We provide brief discussion as follows.

1. Context-Awareness. This involves the ability of

data fusion pipeline to accommodate background

information in addition to the sensor data to de-

velop intelligent fusion methods. The context

may not necessarily directly relate with the sen-

sors, but provide an additional detail that can fur-

ther improve smart home applications. As an ex-

ample, (Dutta and Roy, 2022) fuse environmental

sensors data with 16 types of context data includ-

ing AC status, fan status, door status, heat isola-

tion, location type, area type, and so on, to im-

prove occupancy detection.

2. Reliability. This deals with the reliability of

the results from the application of the data fu-

sion model. Because fusion results can have se-

rious impact in decision making for smart con-

nected home applications, such as activity recog-

nition, diagnoses, emergency predictions, fall de-

tections, ﬁre outbreak predictions, energy con-

sumption predictions, and surveillance, the out-

put of these applications must be reliable to avoid

critical situations. Reliability in data fusion can

be assessed using the established metrics in infor-

mation retrieval. For instance, metrics for classi-

ﬁcation tasks include accuracy, precision, recall,

f-measure, and auc-roc. For clustering tasks, re-

liability can be assessed using metrics like rand

index, Silhouette score, Davies-Bouldin index,

Calinski-Harabasz index, and mutual Informa-

tion. Regression tasks are evaluated using met-

rics like coefﬁcient of determination, root mean

square error, and mean absolute error.

3. Robustness. Data fusion needs to be robust to re-

sist different cyberattacks, such as data injection

and other forms of malicious software exploita-

tion. Since existing solutions utilize the Internet

protocol to transmit raw sensor data to the cloud,

the possibility of data or code injection attacks

cannot be underestimated. Robustness can be as-

sessed based on the level of security offered by the

fusion method.

4. Efﬁciency. This attribute deals with the ability of

data fusion methods to scale irrespective of the

size of the data required for the modeling pro-

cess. Efﬁciency can be assessed in terms of train-

ing time, testing time, communication cost, and

so on.

5. Veriﬁability. Data fusion results should be veriﬁ-

able to ascertain if the contributing sensors actu-

ally improve the ﬁnal fusion results. This can be

assessed by checking the quality of the data col-

lected from each sensor in relation with the smart

connected home application’s objective.

6. Security and Privacy: Data fusion architectures

must ensure that data transmission to the fusion

center is secured and that the data are handled

in conformity with privacy regulation. The re-

sults of the data fusion should also be protected

from unauthorized access or modiﬁcation, and the

availability of the fusion results should not be

compromised.

7. Real-Time. Data fusion methods should de-

liver results to the end-users in real-time. Exist-

ing solutions leverage cloud computing capabil-

ity which offer a near real-time architecture for

IoT applications. Thus, data fusion model needs

to consider real-time delivery of fusion results for

timely decision making.

The above highlighted quality attributes guide the

formulation of our proposed model. HOMEFUS is a

privacy and security-aware model for data fusion tar-

geted toward smart connected homes. Section 5 pro-

vides a detailed discussion of the proposed model.

5 HOMEFUS - THE PROPOSED

MODEL

In this section, we discuss the different components

of the proposed model, and formally guide through its

requirements. HOMEFUS achieves privacy and secu-

rity, aligns with the requirements for data fusion as

previously stated, and includes risk assessment and

mitigation strategies that promote the evaluation of

IoTBDS 2024 - 9th International Conference on Internet of Things, Big Data and Security

136

privacy and security threats directed to the smart con-

nected home (see Fig. 2). The proposed model con-

siders the sensitivity of the data generated and has

the ability to use both intrusive and non-intrusive data

sources since the raw data of the home occupants are

not transmitted to the cloud. It also advocates secure

storage and computation in the cloud to prevent infer-

ence attacks on the local federated machine learning

model that is transmitted from each smart connected

home.

Formally, the proposed approach models a smart

connected home ecosystem as a tuple

(H,U,C, N,A,F,L

,P) where H: smart con-

nected homes, U: users, C: context data, N: con-

nected nodes or devices, A: smart connected home

applications, F: mapping function, L

: local model,

: global model, and P: policy. We provide the de-

tails as follows.

5.1 Smart Connected Homes (H)

This refers to the set of smart connected homes that

have agreed to implement the data fusion pipeline of-

fers by the proposed model. Formally, we identify

a set of smart connected homes H = {h

,..., h

}

where n is the number of homes. To allow for scal-

ability of the model, we assume the value of n is dy-

namic.

5.2 Users (U)

Users represent smart connected home occupants in-

cluding family members and invited guests. A set

of users in home h

, h

∈ H, is denoted as U

,..., u

}, ∀u

∈ U. m is also considered to be

dynamic, because a user u

can join or leave the home

network at any time.

5.3 Context Data (C)

represents context data generated in each h

, ∀h

∈

H at a particular timestamp. Similar to (Dutta and

Roy, 2022), we distinguish two type of context data:

static (e.g location type, area type) and dynamic (e.g

fan status, awake status). Thus, C

= {C

∪C

5.4 Nodes (N)

This represents smart connected devices or nodes. We

identify N

, the set of devices used by user u

in home

, ∀u

∈ U and ∀h

∈ H. These devices have the ca-

pability to sense, actuate, process, and transmit data.

Usually, N

= {C

∪M

∪P

} where C

are connected

devices with ﬁxed location (e.g., a washing machine,

a fridge, etc.); M

are mobile devices (e.g., a smart-

phone, a laptop, etc.), and P

are the processing nodes

(e.g., the edges or the cloud). P

has additional fea-

tures which include storage, processing, and analytic

capability. From a security perspective, they are also

equipped with intrusion detection and prevention sys-

tem (IDPS). Edge nodes are located inside the home

while cloud connection is outside the home. Both

edge and cloud nodes are connected using a smart

home gateway. This gateway is maintained by the

middleware layer. From a security perspective, we

assume that the gateway uses standard protocols with

standard security conﬁgurations. A typical example

is the use of MQTT, which operates using the publish

and subscribe principle and can work with SSL/TLS

encryption. When a sending node publish a topic, the

receiving node can securely subscribe to it. HOME-

FUS leverages homomorphic encryption to secure the

local model to be transmitted via the gateway to the

cloud.

Each device in C

and M

∈ N

can have one or

more sensors S

, j = 1,2, ...,k and i = 1, 2,..., n at-

tached to it. Our proposed model aims to identify the

sensor S

that should be fused with context data C

to improve smart connected home applications and

services. To ensure this, we propose ﬁlter-based sen-

sor selection using information gain and correlation-

based methods. These methods are not computation-

ally expensive compared to the existing state-of-the-

art approaches. Additionally, we chose this method

to take advantage of edge computing and offer a low-

computationally demanding solution.

5.5 Applications (A)

Every activity or behavior of user u

∈ U can signal

a speciﬁc smart connected home application a

∈ A,

where A is denoted as a set of smart home applications

or services, i.e., A = {a

,..., a

ℓ

}. Typical appli-

cations or services include activity recognition (e.g.,

cooking, washing, etc.), ﬁre detection, surveillance,

energy management, remote monitoring, and so on.

5.6 Mapping Function (F)

The role of the mapping function F is to map

sensor S

, j = 1, 2, ..., k to a speciﬁc application

or service a

∈ A. This mapping is denoted as

F({S

,..., S

},a

). The mapping function com-

putes the relevance of each sensor to the IoT applica-

tion or service a

. F relies on both information gain

and correlation-based sensor selection for the map-

ping.

HOMEFUS: A Privacy and Security-Aware Model for IoT Data Fusion in Smart Connected Homes

137

Figure 2: Proposed model for data fusion in smart connected home.

Formally, given ℓ number of applications, infor-

mation gain (IG) can be obtained from entropy as fol-

lows.

Entropy(A) = −

ℓ

∑

i=1

p(a

)log

(p(a

)) (1)

where p(a

) is the probability of A extracted accord-

ing to application a

. The contribution of each sensor

is estimated according to Eqn. 2.

(A,S

) = Entropy(A) −

∑

v∈S

|A|

.Entropy(S

) (2)

where F

(A,S

) is a mapping function that com-

putes the usefulness of sensor S

for applications A

based on information theoretic. This quantiﬁes the in-

formation gained by this sensor. The higher the value

of F

(A,S

), the more useful the sensor S

is.

Similarly, a merit score, F

can be obtained for

the sensors using correlation-based sensor selection

method as in Eqn. 3.

({S

,..., S

},A) =

d + d(d − 1)r

(3)

where d is the number of sensors in the subset, r

the mean of application-to-sensor relevance correla-

tion, and

is the mean of sensor-to-sensor relevance

correlation. The subset of sensors with the highest

merit is selected as the output of the correlation-based

sensor selection approach.

We then deﬁne F

IGM

= {F

(.)∪ F

(.)}, with any

duplicate removed, which represent the sensors se-

lected by information gain and correlation sensors se-

lection approaches that are considered in the proposed

model. F

IGM

is then merged with the context data C

i.e F

= {F

IGM

∪C

}, where F

denote the data used

to develop the local federated learning model on the

edge of each home, h

∈ H.

5.7 Federated Learning

The proposed model employs federated learning to

enhance privacy of home occupants. In this setting,

the raw data of home users U are not transmitted

directly to the cloud, but rather the learned model

trained from the individual home hi ∈ H who have

subscribed to participate in the ecosystem. This local

model is encrypted using homomorphic encryption.

Homomorphic encryption allows computation on en-

crypted data in the cloud. In our proposed context,

federated learning allows multiple homes to collabo-

ratively train a machine learning model without nec-

essarily sharing their data with each other. For the

machine learning model, we propose gradient boost-

ing methods due to the following reasons: accuracy,

train faster especially on large datasets, and capable

of handling noisy data (Mwiti, 2023). In practice, a

federated XGBoost algorithm can serve this purpose

(IBM, 2023).

IoTBDS 2024 - 9th International Conference on Internet of Things, Big Data and Security

138

Formally, let L

denotes the local model trained

at the edge node and G

represents the global model

updated at the cloud node. The learning can be esti-

mated as given in Eqn. 4.

= argmin

∑

i=1

E(F

s j

,a j) ∼ L

[ℓ(w, F

s j

,a j)]

(4)

where w is the model parameters, n is the number of

homes, L

is the probability distribution of data at

home h

, ℓ(.) is the loss function, and E(.) is the ex-

pected value. The goal of the model is to obtain the

global parameter G

that minimizes the loss function.

5.8 Policy (P)

Policy regulates the different security and privacy as-

pects of the model. These are the rules that can be

implemented to ensure security and privacy. Policy

is denoted as P = {p

, p

,..., p

} where p

∈ P may

include fusion policy, data transfer policy, data reten-

tion policy, encryption policy, key management pol-

icy, storage policy, and model update policy. For in-

stance, model update policy may include a rule that

specify when a local model should be updated. Con-

sidering this policy, a timestamp attribute can be used

to automate the update process. The fusion policy

will include a rule that ensures only the authenticated

sensor nodes are allowed to connect. Other policies,

∈ P, will also have speciﬁc rules that guide the

model in a collaborative manner.

6 EVALUATION AND RISK

ASSESSMENT

The proposed model is formulated taking into con-

sideration the requirements for data fusion. The fol-

lowing privacy and security threats are identiﬁed. We

brieﬂy highlight how they are addressed in the pro-

posed model.

6.1 Privacy Threats

• Linkage. An attacker wants to reveal information

that is not disclosed by home occupants by linking

data from different sources. Since raw data of the

users are not transmitted, this threat is minimized.

• Localization and Tracking. An attacker wants to

record the location of home occupants and track

their movements. The proposed model preserves

users privacy and the IoT devices are conﬁgured

to transmit data to the edge nodes, and in doing

so; this type of attack is prevented.

• Model Inversion. An attacker wants to determine

if a particular home or entity has been used to de-

velop the local or global model. The countermea-

sure here is that the communication between the

edge and the cloud is encrypted, making it difﬁ-

cult to compromise the model.

• Proﬁling. A hacker wants to collect and correlate

data to generate new data about home occupant.

The main countermeasure is that raw data trans-

mission is not permitted by the data transfer pol-

icy.

• Inventory Attack. An attacker wants to gain unau-

thorized access to the home network and gather

occupant data from an edge node. Since the edge

node is equipped with IDPS, this will prevent

unauthorized data collection.

6.2 Security Threats

• Conﬁdentiality. A hacker wants to have the

knowledge about the data in transit and those at

storage to compromise data integrity. The coun-

termeasure here is that the policy governing the

model does not permit raw data transfer between

the gateway and the cloud. The cloud provides se-

cured storage and computation for the model re-

sults, making it difﬁcult for attacker to succeed

in this respect. Authentication and access control

mechanisms are also provided in the cloud and the

edge nodes, which also strengthen conﬁdentiality.

• Data Integrity and Data Poisoning. An attacker

wishes to modify the data in transit and to inject

fake data to compromise data integrity. The possi-

bility of this attack is limited since authentication

and access control mechanisms are provided. In

addition, only the model trained is transmitted in

encrypted form, which also reinforces the means

to ensure data integrity.

• Eavesdropping. This is also known as snifﬁng or

snooping. In this type of attack, the hacker re-

lies on unsecured network communications to ac-

cess data in transit between nodes or devices. The

possibility of this attack is limited since there is a

secured connection between the gateway and the

cloud for the model transfer.

• Denial of Service. A hacker wants to compromise

availability of the data fusion pipeline by ﬂood-

ing the network with superﬂuous requests. This

attack is prevented by the IDPS that is running on

the edge and the cloud. In addition, the middle-

ware also ensures that only the secured connec-

tions and sessions are routed, and that requests are

time-bound.

HOMEFUS: A Privacy and Security-Aware Model for IoT Data Fusion in Smart Connected Homes

139

7 CONCLUSION

In this paper, a privacy and security-aware model

for smart connected home applications is proposed.

It advocates privacy and security for IoT data fu-

sion in smart pervasive living spaces where a lot of

personal data is generated, stored, and distributed.

In the model, the requirements for efﬁcient data fu-

sion pipeline are considered, and federated learning

to protect home occupants’ data and improve predic-

tive analysis are adopted. Edge nodes are considered

for local model training and deployment, and a se-

cure connection is established between the edge and

the cloud. We show that the proposed model meets

the requirements for efﬁcient data fusion and that it

can be applied to a variety of smart connected home

applications and services. Future work will consider

empirical analysis of the performance of the proposed

model, considering its different components.

ACKNOWLEDGMENT

This work was partially funded by the Knowledge

Foundation (Stiftelsen f

or kunskaps- och kompeten-

sutveckling – KK-stiftelsen) via the Synergy project

Intelligent and Trustworthy IoT Systems (Grant num-

ber 20220087).

REFERENCES

Adewole, K. S. and Torra, V. (2022a). Dftmicroagg: a dual-

level anonymization algorithm for smart grid data. In-

ternational Journal of Information Security, pages 1–

23.

Adewole, K. S. and Torra, V. (2022b). Privacy is-

sues in smart grid data: From energy disaggregation

to disclosure risk. In International Conference on

Database and Expert Systems Applications, pages 71–

84. Springer.

ıvodji, U. M., Gambs, S., and Martin, A. (2019). Iot-

ﬂa: A secured and privacy-preserving smart home ar-

chitecture implementing federated learning. In 2019

IEEE security and privacy workshops (SPW), pages

175–180. IEEE.

Bugeja, J., Vogel, B., Jacobsson, A., and Varshney, R.

(2019). Iotsm: an end-to-end security model for iot

ecosystems. In 2019 IEEE International Conference

on Pervasive Computing and Communications Work-

shops (PerCom Workshops), pages 267–272. IEEE.

Chaaraoui, A. A., Padilla-L

opez, J. R., Ferr

andez-Pastor,

F. J., Nieto-Hidalgo, M., and Fl

orez-Revuelta, F.

(2014). A vision-based system for intelligent monitor-

ing: human behaviour analysis and privacy by context.

Sensors, 14(5):8895–8925.

Chahuara, P., Portet, F., and Vacher, M. (2013). Making

context aware decision from uncertain information in

a smart home: A markov logic network approach.

In Ambient Intelligence: 4th International Joint Con-

ference, AmI 2013, Dublin, Ireland, December 3-5,

2013. Proceedings 4, pages 78–93. Springer.

Chimamiwa, G., Alirezaie, M., Pecora, F., and Loutﬁ, A.

(2021). Multi-sensor dataset of human activities in a

smart home environment. Data in Brief, 34:106632.

Ding, W., Jing, X., Yan, Z., and Yang, L. T. (2019). A sur-

vey on data fusion in internet of things: Towards se-

cure and privacy-preserving fusion. Information Fu-

sion, 51:129–144.

Dutta, J. and Roy, S. (2022). Occupancysense: Context-

based indoor occupancy detection & prediction using

catboost model. Applied Soft Computing, 119:108536.

Gartner Inc. (2017). Gartner Says 8.4 Billion Connected

”Things” Will Be in Use in 2017, Up 31 Percent From

2016. https://www.gartner.com/en/newsroom/press-r

eleases/2017-02-07-gartner-says-8-billion-connect

ed-things-will-be-in-use-in-2017-up-31-percent-fro

m-2016. Accessed: 5th June, 2023.

IBM (2023). Federated learning xgboost tutorial for ui. ht

tps://www.ibm.com/docs/en/cloud-paks/cp-data/4.

6.x?topic=samples-xgboost-tutorial. Accessed: 5th

June, 2023.

Kommey, B. (2022). Automatic ceiling fan control using

temperature and room occupancy. JITCE (Journal of

Information Technology and Computer Engineering),

6(01):1–7.

Mena, A. R., Ceballos, H. G., and Alvarado-Uribe, J.

(2022). Measuring indoor occupancy through envi-

ronmental sensors: A systematic review on sensor de-

ployment. Sensors, 22(10):3770.

Monti, L., Tse, R., Tang, S.-K., Mirri, S., Delnevo, G.,

Maniezzo, V., and Salomoni, P. (2022). Edge-based

transfer learning for classroom occupancy detection in

a smart campus context. Sensors, 22(10):3692.

Mwiti, D. (2023). Gradient boosted decision trees [guide]:

a conceptual explanation. https://neptune.ai/blog/gr

adient-boosted-decision-trees-guide. Accessed: 5th

June, 2023.

Pathmabandu, C., Grundy, J., Chhetri, M. B., and Baig, Z.

(2023). Privacy for iot: Informed consent manage-

ment in smart buildings. Future Generation Computer

Systems, 145:367–383.

Rahman, A. F. A., Daud, M., and Mohamad, M. Z. (2016).

Securing sensor to cloud ecosystem using internet of

things (iot) security framework. In Proceedings of

the International Conference on Internet of things and

Cloud Computing, pages 1–5.

Sayed, A. N., Bensaali, F., Himeur, Y., and Houchati, M.

(2023). Edge-based real-time occupancy detection

system through a non-intrusive sensing system. En-

ergies, 16(5):2388.

IoTBDS 2024 - 9th International Conference on Internet of Things, Big Data and Security

140