CDAC: Content-Driven Access Control Architecture for Smart Farms

Ghadeer I. Yassin

1 a

and Lakshmish M. Ramaswamy

2 b

School of Computing, Binghamton University, Binghamton, NY, U.S.A.

School of Computing, University of Georgia, Athens, GA, U.S.A.

Keywords:

Smart Farm, Internet of Things, Edge Machine Learning, Image Classiﬁcation, MobileNet, Access Control.

Abstract:

One of the smart farms’ pivotal components involves leveraging vast quantities of imagery data to inform

decision-making and improve farm outcomes. With the increasing integration of image data in smart farms,

ensuring secure and efﬁcient access to these data sets is crucial. This paper proposes a novel Content-Driven

Access Control (CDAC) architecture designed speciﬁcally for smart farming environments, where access re-

quests to image data are evaluated based on the visual content of the images. The CDAC architecture employs

a novel technique to assess the relevance of access requests to speciﬁc image contents by enriching access con-

trol requests with useful image content information with the help of an edge machine learning classiﬁcation

model that provides a fast and small-weight solution to classify images near their source in the smart farm.

This approach goes beyond traditional access control methods by considering the information within images,

allowing for more granular and content-aware permissions. To validate the effectiveness of the CDAC archi-

tecture, a series of experiments were conducted using a dataset of agricultural images. Results demonstrate

that the proposed architecture is a valuable solution for regulating access to smart farm images based on the

visual content of the images. Additionally, the architecture is proven suitable for deployment on smart farm

edge devices.

1 INTRODUCTION

Smart farms represent a transformative paradigm in

modern agriculture, where the integration of cutting-

edge technologies has given rise to a wealth of col-

lected images that redeﬁne how we understand and

manage farm and agricultural operations. Cameras,

drones, and sensors deployed across smart farms cap-

ture diverse visual data, ranging from real-time snap-

shots of crops and livestock to high-resolution aerial

views of entire ﬁelds. These images serve as in-

valuable datasets, providing insights into crop health,

environmental conditions, and the overall status of

the farm. The collected imagery forms the founda-

tion for informed decision-making, enabling farmers

and stakeholders to monitor, analyze, and respond

promptly to dynamic agricultural scenarios. However,

the sheer volume of image data generated necessitates

sophisticated approaches to image management, in-

cluding restriction to image retrieval based on their

contents.

The current landscape of access controls in smart

https://orcid.org/0000-0001-5135-7622

https://orcid.org/0000-0002-4567-4186

communities, and speciﬁcally in smart farming, high-

lights signiﬁcant shortcomings, particularly in the

domain of regulating access to image data. Tradi-

tional access control mechanisms predominantly rely

on broad resource attributes such as image source or

given title, overlooking the nuanced nature of image

content.

In implementing robust security measures for gov-

erning access to images based on their contents within

the realm of smart farming, the integration of image

classiﬁcation models is paramount. However, Image

classiﬁcation for access control purposes faces multi-

ple challenges. Those challenges include the security

concerns raised by transmitting sensitive farm images

to the cloud-based systems where image classiﬁcation

models can be deployed and the impracticality of de-

ploying these models on smart farm edge limited ca-

pability devices as a counter solution. Conversely,

there’s a pressing demand for instantaneous classi-

ﬁcation of farm-collected images to expedite access

control decisions. Moreover, banking solely on im-

age classiﬁcation results for access control decisions

proves inefﬁcient due to the inability to ensure a 100%

accuracy rate from the classiﬁcation models.

Yassin, G. I. and Ramaswamy, L. M.

CDAC: Content-Driven Access Control Architecture for Smart Farms.

DOI: 10.5220/0013372000003899

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

In Proceedings of the 11th International Conference on Information Systems Security and Privacy (ICISSP 2025) - Volume 1, pages 83-94

ISBN: 978-989-758-735-1; ISSN: 2184-4356

To tackle these limitations, our paper introduces a

novel smart farm content-driven access control archi-

tecture composed of an edge machine learning clas-

siﬁer that provides a fast and small-weight solution

to classify images near their source in the smart farm

and employs a semantic attribute-based access control

mechanism that enriches access control requests with

useful image content data to reduce undesired access

to farm images. Our research contribution includes:

• Identifying the limitations of traditional access

control methods in regulating accessibility to im-

age resources based on their visual contents,

paving the way for more granular and content-

aware access control solutions.

• Creating a novel content-driven access control ar-

chitecture for the smart farm domain.

• Introducing an instant classiﬁcation technique of

newly captured farm images enabling swift access

control decisions while considering smart farm

images’ different contents security levels.

• Assessing the effectiveness of the suggested ar-

chitecture on edge devices to better align with the

unique requirements of smart farming.

The paper is structured as follows. In Section

2, we draw the motivation behind this research and

in Section 3 we lay out important background infor-

mation. While, in Section 4, we provide a detailed

overview of the proposed architecture. In Section 5

we describe the prototype implementation of the ar-

chitecture as well as the different evaluation experi-

ments. In Section 6 we discuss our evaluation results

and in section 7, we discuss important related works.

Finally, in Section 8, we draw our conclusions.

2 MOTIVATION

Our research is motivated by the need to address

the limitations of existing access control solutions in

managing image data access within smart farming

systems. In these systems, a wide range of stakehold-

ers with different roles or specialties might require

them to access and examine different imagery data

related to their work. Conversely, it is imperative to

enforce restrictions to prevent these stakeholders from

accessing or reviewing imagery data outside the scope

of their expertise. Take, for instance, an agronomist

dedicated to fruit cultivation within the framework

of a smart farming system. The agronomist’s proﬁ-

ciency centers on the meticulous evaluation of fruit

crops, requiring the analysis of visual data to assess

factors such as health, growth, and potential chal-

lenges speciﬁc to various fruits. In this context, the

agronomist necessitates exclusive access to images

featuring fruits. However, their role does not extend to

the examination of all-encompassing smart farm im-

agery, which may include images of other stakehold-

ers working inside the smart farm, agricultural equip-

ment visuals, or livestock monitoring images. These

broader datasets might encompass security-sensitive

content not directly relevant to the agronomist’s re-

sponsibilities.

To address the agronomist’s access needs and

safeguard other farm images, Access control poli-

cies should be meticulously formulated to grant the

agronomist exclusive access to images featuring fruits

while simultaneously restricting their access to im-

ages containing other contents. The access control

solution not only enhances the agronomist’s ability to

focus on critical aspects of fruit health but also safe-

guards images unrelated to their specialization, con-

tributing to the effectiveness of image management

practices within the smart farming system.

However, The existing access control solutions in

smart communities, particularly in smart farming, re-

veal substantial limitations, particularly regarding the

management of image data access. Traditional access

controls mainly focus on broad resource attributes

such as image IRI, source, or provided title, failing

to consider the nuanced nature of image content. This

lack of granularity hampers control over stakehold-

ers’ access to speciﬁc content within the vast array of

image data generated in smart farming. To overcome

these limitations, a more sophisticated access control

approach is needed for smart farming systems to en-

sure that stakeholders have precise access to the im-

age data relevant to their expertise while maintaining

security by limiting accessibility to other contents.

Efﬁcient access control solutions within smart

farms should identify images based on their contents

and meticulous access control policies should be for-

mulated to exclusively grant the stakeholders access

to images bearing the speciﬁed content while concur-

rently imposing restrictions on their access to unre-

lated images. This bespoke approach fortiﬁes security

by preventing access to irrelevant or security-sensitive

imagery resulting in an advancement in both the efﬁ-

ciency and security of image management practices

within the smart farm. On the other hand, estab-

lishing robust security measures to manage access to

images based on their content within the context of

smart farming, requires the incorporation of machine

learning models for image classiﬁcation. While con-

ventional methods, involving the processing of im-

age data in cloud-based systems for machine learn-

ing tasks, exhibit effectiveness, they raise signiﬁcant

security concerns due to the transmission of sizable

ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy

volumes of potentially sensitive visual information.

This underscores the vital transition towards de-

ploying such models at the edge of the smart farm-

ing network, closer to the data sources. However, the

computational demands inherent in machine learn-

ing models encompass intensive computations, result-

ing in models that are computationally intensive and

reliant on high-performance computing resources.

These challenges are particularly pronounced when

extending deep networks to edge devices in smart

farming, where constraints such as limited battery ca-

pacity and memory further compound the intricacies.

Hence, there is an imperative need to develop

models tailored to the constrained computational re-

sources of edge devices in smart farms, recognizing

the unique limitations of these environments. This

strategic approach ensures not only enhanced security

but also optimized performance in real-world applica-

tions within the context of smart farming.

Additionally, it’s imperative to recognize that

achieving 100% classiﬁcation accuracy is unrealis-

tic. Consequently, an access control architecture may

face the risk of making decisions based on inaccu-

rate classiﬁcation results. This risk becomes particu-

larly severe when granting access to images contain-

ing highly sensitive visual content. Conversely, deny-

ing access due to incorrect classiﬁcation, while less

severe, can still disrupt essential tasks. Therefore,

it’s essential for the access control architecture to ad-

dress the imperfections of image classiﬁcation results

and implement measures to mitigate these risks effec-

tively.

Furthermore, smart farms generate vast amounts

of farm image data daily, which serve as the founda-

tion for swift decision-making in smart farming op-

erations. Therefore, the access control architecture

should not restrict access solely to pre-classiﬁed im-

ages. Instead, it should integrate instant image clas-

siﬁcation techniques to ensure that access control re-

quests are promptly evaluated based on image con-

tents.

3 BACKGROUND INFORMATION

Enabling an access control solution to assess access

requests to images based on their content requires

comprehensive information about the image content.

These details can be generated with the help of edge-

suitable image classiﬁcation models. Subsequently,

enriching them within access control requests to serve

as speciﬁcs for the targeted resource, namely the farm

image. Consequently, the access control model can

accurately evaluate the request based on the informa-

tion about the image content.

While various solutions for semantically enrich-

ing access requests with metadata about targeted re-

sources exist, tailored approaches are necessary for

smart farms due to their distinctive characteristics

(Yassin and Ramaswamy, 2022). Hence, we build

upon the semantically enriched access control archi-

tecture proposed in (Yassin and Ramaswamy, 2023)

and incorporate within it a quantized Mobilnet model

to perform image classiﬁcation tasks on smart farm

edge devices.

In this section, we brieﬂy describe the important

components of our proposed solution.

3.1 Semantically-Enriched Access

Control

Figure 1: RE Component.

In the semantically enriched access control ar-

chitecture (Yassin and Ramaswamy, 2023), the tra-

ditional ABAC architecture was modiﬁed to include

a new RE component with two modules namely at-

tribute processor and semantic reasoner as described

in Figure 1. The attribute processor module extracts

information from the access requests and forwards

them to the semantic reasoner. Utilizing a smart farm-

speciﬁc ontology, the semantic reasoner deduces sup-

plementary attributes about the request entities, which

are then returned to the attribute processor for incor-

poration into the request. This process empowers the

Policy Decision Point (PDP) to make well-informed

access control decisions.

3.2 Mobilnet Models

Mobilenets are machine learning classiﬁcation mod-

els that are proven suitable for edge devices. It is

a family of neural network architectures speciﬁcally

crafted for efﬁcient computation on mobile devices

with constrained computing power and memory re-

sources (Howard, 2017). It strategically employs

CDAC: Content-Driven Access Control Architecture for Smart Farms

depth-wise separable convolutions, breaking down

the standard convolution into depth-wise and point-

wise convolutions. This approach signiﬁcantly re-

duces the number of parameters and computations re-

quired for image recognition and classiﬁcation tasks,

enhancing overall efﬁciency. Furthermore, the Mo-

bileNet V2 model incorporates a technique known

as ’bottlenecking’ to further decrease computational

complexity by compressing the input feature map be-

fore processing it with convolutional layers.

MobileNet models’ architecture design, with its

emphasis on depth-wise separable convolutions, aims

to address the challenges of deploying deep learn-

ing models on mobile devices with limited resources.

This makes it well-suited for smart farm applications.

In this paper, MobileNet-V1 and MobileNet-V2 were

selected for our use case.

3.3 Model Quantization

Quantization is a commonly applied technique to net-

works, aiming to reduce the number of bits needed

to represent weights without compromising accuracy.

One prevalent method is the utilization of lower preci-

sion formats such as 16-bit ﬂoating-point (FP16) and

32-bit ﬂoating-point (FP32) for representing weights

during the quantization process (Cheng et al., 2017).

This allows for a reduction in the memory footprint

and computational demands of the model, making it

more suitable for deployment on resource-constrained

edge devices and mobile applications.

In our proposed architecture, we have employed

the quantization approach as a key strategy for model

compression. The optimal representation for our spe-

ciﬁc use case has been determined through an empir-

ical analysis conducted directly on edge devices.

4 PROPOSED ARCHITECTURE

The pipeline of the proposed architecture is illustrated

in Figure 2. Our proposed architecture employs a

multi-path approach to enhance access control deci-

sions. When the system receives an access request

for image resources, it initiates a dynamic process

to determine whether the targeted resource is a pre-

classiﬁed image or a new, unclassiﬁed image.

When an unclassiﬁed image is received, it under-

goes classiﬁcation using an edge-suitable image clas-

siﬁcation model. This model analyzes the visual con-

tent, extracting information that directly enhances the

real-time access request. Simultaneously, the classi-

ﬁcation details, such as the Image IRI (Internation-

alized Resource Identiﬁer) and the probability distri-

Figure 2: Proposed Pipeline.

bution of image content classes, seamlessly integrate

into the smart farm ontology, creating a structured

representation of the image content as described in

Figure 3. This structured representation lays the foun-

dation for subsequent access requests.

In cases where the image is pre-classiﬁed, the sys-

tem taps into the semantic richness stored in the on-

tology to deduce image content information. This in-

formation is then utilized to enrich the access request.

Figure 3: Image Classiﬁcation Pipeline.

In both scenarios, the access requests are en-

riched with relevant image content information. This

augmentation empowers the system to make well-

informed decisions about granting or revoking acces-

sibility to images based on their contents rather than

their general attributes (Source, Title, etc). This ap-

proach allows the system to gain a deeper understand-

ing of the visual content, enabling it to make access

decisions based on contextual insights derived from

both pre-classiﬁed and dynamically classiﬁed images.

ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy

5 ARCHITECTURE SETUP

5.1 Dataset

The dataset selected for the model training and eval-

uation process is the Kaggle ”Plants Type Dataset”

which presents a robust collection of 30,000 high-

resolution plant images. With a meticulous curation

of 1,000 images per class, the dataset spans across 30

distinct plant classes and encompasses seven diverse

plant types, including crops, fruits, vegetables, and

herbs.

5.1.1 Dataset Split

To promote a comprehensive evaluation of model

performance, The dataset was partitioned into train-

ing, validation, and testing subsets. The training set,

which constitutes 70% of the data, is employed to

train the model, enabling it to learn and generalize

from diverse examples. Simultaneously, the valida-

tion set, comprising 15% of the dataset, is reserved for

assessing the model’s performance on unseen data,

providing insights into its robustness. While the re-

maining 15% of the data were reserved for the phase

of model testing.

5.1.2 Image Preprocessing

Before model training, a comprehensive preprocess-

ing pipeline was implemented to optimize the dataset

for enhanced neural network learning and generaliza-

tion. Pixel values of images were normalized to a

standardized range of [0, 1] to enhance the model’s

convergence during training. Simultaneously, all im-

ages were resized to uniform 224x224 dimensions,

establishing a standardized input size to ﬁt the in-

put shape for the subsequent convolutional neural net-

work input. Furthermore, images’ randomness was

introduced through shufﬂing images during training

to enhance the model’s adaptability by exposing it to

varied patterns and features.

5.2 Experiments

For our experiments, we employed transfer learn-

ing techniques which involve retaining knowledge ac-

quired while addressing one problem and applying it

to solve a new, related problem enabling the model to

adapt more swiftly to new data, a process that would

be less efﬁcient if started from scratch.

We utilized MobileNet models with pre-trained

weights from ImageNet, a dataset containing millions

of well-organized images, accessible globally to re-

searchers (Deng et al., 2009). To enhance the models’

performance, ﬁne-tuning was conducted in this study.

Fine-tuning involves carefully adjusting the model’s

parameters to better respond to speciﬁc information.

During this process, a trained model or a segment of it

is unfrozen, and the training occurs anew on new data,

utilizing a reduced learning rate. This slight modiﬁca-

tion of the already learned weights results in enhanced

performance.

The primary method employed for ﬁne-tuning in

this investigation involves removing the last fully con-

nected layer from selected pre-trained CNN models

and replacing it with a new fully connected layer or

layers, matching the number of classes in our dataset.

In our case, we employed 30 classes due to the spe-

ciﬁc requirements of our dataset. It’s important to

note that we employed two strategies: one where

the MobileNet models were used as feature-extractors

with all layers frozen, and another with ﬁne-tuning,

where some layers were unfrozen. This dual approach

allowed us to leverage the general features learned by

MobileNet models while also adapting the model to

the nuances of our speciﬁc classiﬁcation task.

5.2.1 Hyperparameters Selection

The models were evaluated with the goal of classify-

ing the 30 different classes found in the data set. A

training duration of 25 epochs was utilized for calcu-

lating model metrics, and the choice of the number of

epochs and learning rate parameters was made empir-

ically. Past research has demonstrated promising out-

comes with the use of Adam as an optimizer (Bera and

Shrivastava, 2020), inﬂuencing our decision to adopt

Adam as the optimizer for our models. Additionally,

a batch size of 32 and a learning rate of 0.0001 were

speciﬁed. The complete set of hyperparameters for all

transfer learning models is outlined in Table 1.

Table 1: Hyperparamters For All Transfer Models.

Parameter Value

Max Epochs 25

Min Patch size 32

Optimizer Adam Optimizer

Learning Rate 0.0001

Loss Function Categorical Crossentropy

5.2.2 Model Quantization

To implement model quantization, the proposed

framework utilizes the Tensorﬂow Lite Library, em-

ploying various quantization parameters to evaluate

the impact on model size and performance. Speciﬁ-

cally, we represented the selected model using both

CDAC: Content-Driven Access Control Architecture for Smart Farms

16-bit ﬂoating-point precision (FP16) and 32-bit

ﬂoating-point precision (FP32). Subsequently, a com-

prehensive evaluation was conducted on both repre-

sentations to assess their effectiveness in balancing

model size reduction and preserving accuracy. .

5.3 Smart Farm Ontology

Building upon the classiﬁcation results from our

model, we extended the ontology detailed in (Yassin

and Ramaswamy, 2023) to encompass information re-

garding classiﬁed images. Using the RDFLib Library,

our code dynamically loads the ontology and system-

atically incorporates image instances into the graph.

Each image instance from the dataset is assigned

to the ontology class ’cropImageResource’ and is

linked to class probabilities, reﬂecting the likelihood

of the image belonging to one of the 30 possible

categories, along with associated conﬁdence scores.

Additionally, we group conﬁdence scores for major

classes based on the classiﬁcation of 30 classes as fol-

lows:

• Fruits: waterapple, pineapple, pomelo, guava,

mango, papaya, orange, banana, cantaloupe, wa-

termelon, coconut, bilimbi.

• Vegetables: soybeans, pepper chili, spinach, shal-

lot, sweet potatoes, ginger, kale, long beans, egg-

plant, cucumber.

• Crops: tobacco, melon, paddy, cassava, corn.

• Herbs: galangal, curcuma, aloe vera.

For instance, consider an image classiﬁcation re-

sults as: (Waterapple, 25%), (Mango, 30%), (Pepper,

20%), (Orange, 15%),( Bilimbi, 10%). Therfore, the

image classiﬁcation for the broader categories would

be: (Fruits,80%),(Vegetables:20%).

The resulting RDF triples, encapsulating these as-

sociations, are serialized and seamlessly integrated

back into the ontology, enriching it with pertinent

image-related data and enhancing the knowledge rep-

resentation within the Smart Farm ontology frame-

work. This augmentation facilitates a more compre-

hensive understanding of the images by encompass-

ing valuable insights derived from image classiﬁca-

tion results.

5.4 Request Enricher- RE

The RE component, outlined in Section 3, plays a piv-

otal role in our content-driven access control architec-

ture and was modiﬁed as shown in Figure 4. Within

the RE component, the Attribute Processor module

extracts crucial attributes from the access control re-

quest, notably the targeted resource, which in this

context is a Smart Farm Collected Image. These

attributes are then forwarded to either the Semantic

Reasoner module or the Image classiﬁcation model

directly and infer information about the image con-

tent, speciﬁcally its class and conﬁdence score. The

inferred details are subsequently sent back to the At-

tribute Processor.

Figure 4: RE component changes.

5.4.1 Class Speciﬁc Threshold

Building upon the results of image classiﬁcation, a

single image may be assigned to multiple classes,

each associated with distinct conﬁdence scores. In

this context, we consider a security-sensitivity level

assigned to images corresponding to different classes

to mitigate the imperfections of image classiﬁcation

results. For instance, images of herbs might carry

a higher security sensitivity compared to images of

fruits. Consequently, when the additional attributes

are received, the Attribute Processor component un-

dertakes a comparison between the image contents

attributes and class-speciﬁc thresholds. If the conﬁ-

dence score surpasses the designated threshold, the

image content attribute is deemed suitable for request

enrichment; otherwise, it is disregarded.

5.4.2 Request Update

Upon receiving the additional attributes and compar-

ing them to the speciﬁed sensitivity threshold, the At-

tribute Processor appends the selected attributes to the

original request. The enriched request, now contain-

ing information about the targeted image content, is

then sent to the Policy Decision Point (PDP) for eval-

uation. This content-driven approach enhances the

access control decisions by incorporating detailed in-

sights into the nature of the requested resource, In this

context, the Image content.

ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy

5.5 Policy Adminstration

As detailed in Table 2, the access control policy set

incorporates nuanced regulations governing access

to resources (Smart Farm collected Images). This

involves regulating access based on the content of

the images and the subject’s specialization, allowing

them to access image contents in alignment with their

expertise. The policies introduce the criteria con-

sidering the content of the images, and subject spe-

cializations. These policies are designed to facili-

tate a more granular and content-aware access control

mechanism, ensuring that individuals can only access

image content relevant to their speciﬁc area of exper-

tise.

6 EVALAUATION RESULTS

For evaluation purposes, The initial phases of evalua-

tions targeted different models’ evaluations, including

training, validation, and testing, were performed on

Google Colab. This cloud-based platform provided

the computational resources necessary for the train-

ing phase, ensuring efﬁcient model development.

The subsequent evaluation criteria were thor-

oughly tested on the Raspberry Pi 4 Model B to as-

sess the performance and feasibility of our proposed

system in a real-world, edge computing environment.

The following evaluation criteria were tested:

• Model Performance: This assessment aims to

identify the most effective transfer learning model

conﬁguration for the task by evaluating its per-

formance across training, validation, and testing

phases. The purpose is to discern which model

excels in terms of accuracy, generalization, and

effectiveness for the given application. This eval-

uation provides insights crucial for selecting the

optimal model for deployment.

• Model Evaluation on Edge: Assessment of the

performance of the selected transfer learning im-

age classiﬁcation model on a Raspberry Pi device.

The evaluation includes inference time, quantiza-

tion effects on accuracy, and model size consider-

ations, aiming to identify the most suitable conﬁg-

uration for efﬁcient deployment on edge devices.

• Architecture Ability to Enrich Access Re-

quests: Assessment of the capability of the pro-

posed architecture in handling access requests to

images through enriching access requests with

image content details.

• Threshold Adjustment Effect: Assessment of

the effect of adjusting class-speciﬁc threshold in

mitigating the effect of making access request de-

cisions based on inaccurate image classiﬁcation

results.

• Average Request Evaluation Time: Assessment

of the time required on average for a single request

to be processed on Edge device. The evaluation

includes image content inference from both smart

farm ontology and the image classiﬁcation model.

6.1 Model Performance

Figures 5, 6, 7, and 8 illustrate the Accuracy and Loss

graphs for each model during both training and vali-

dation phases. The training graphs depict the evolu-

tion of network accuracy at 5-epoch intervals, show-

casing the model’s progress during training. Addi-

tionally, the ﬁgures display cross-entropy loss, pro-

viding insights into the models’ overall classiﬁcation

accuracy. The discrepancy between training and val-

idation accuracy is minimal, indicating robust gen-

eralization on the training dataset and effective per-

formance on the validation dataset. Finetuned Mo-

bilNet V1 and Finetuned MobilNet V2 have the best

comparable outcomes, with Finetuned MobilNet V2

showing a slightly superior performance. Notably, all

models demonstrate validation accuracy values sur-

passing their training accuracy, and the ﬁgures sug-

gest an absence of overﬁtting, with the models con-

verging within 25 epochs.

The analysis presented in Table 3 emphasizes that

the models achieved their highest testing accuracy

when subjected to ﬁne-tuning, notably with Finetuned

MobilNet V1 reaching 85% accuracy and Finetuned

MobilNet V2 achieving 86%, showcasing a slight su-

periority in accuracy.

This superior performance of Finetuned Mobil-

Net V2 can be attributed to its underlying architec-

ture, characterized by inverted residuals, linear bot-

tlenecks, and shortcut connections. When ﬁne-tuned,

this architecture exhibited enhanced capabilities for

the speciﬁc task at hand. The ﬁne-tuning process

allowed the model to adapt and specialize, resulting

in improved accuracy compared to Finetuned Mobil-

Net V1 and other conﬁgurations.

6.2 Model Evaluation on Edge

The optimal model chosen for deployment on Rasp-

berry Pi edge device was MobileNet-V2 FineTuned

based on its relatively superior performance. Addi-

tionally, to assess the deployment’s performance, two

different quantization techniques, FP16 and FP32,

were applied to the model. Table 4 describes the

model performance on Edge.

CDAC: Content-Driven Access Control Architecture for Smart Farms

Table 2: Smaple Polices from the Policy Set.

# Subject Subject Property Action Resource Resource Property

1 Regular Worker Speciality: Paddy Read Image Resources Content: Paddy

2 Agronomist Speciality: Crops Read Image Resources Content: Crops

4 Researcher Speciality: Cross-Category-Analysis Read Image Resources Content: Crops, Herbs

Table 3: Different Models Evaluation Metrics.

Training Validation Testing

Model Acc. Prec. Rec. F1 Acc. Prec. Rec. F1 Acc. Prec. Rec. F1

FeatureExtractor MobileNet-v1 0.86 0.92 0.82 0.8678 0.87 0.92 0.83 0.8741 0.77 0.84 0.74 0.7856

Finetuned MobileNet-v1 0.96 0.96 0.96 0.9583 0.97 0.97 0.97 0.9705 0.85 0.86 0.85 0.8556

FeatureExtractor MobileNet-v2 0.89 0.95 0.82 0.8797 0.89 0.95 0.83 0.8798 0.80 0.88 0.74 0.8020

Finetuned MobileNet-v2 0.97 0.96 0.96 0.9575 0.97 0.97 0.97 0.9696 0.86 0.87 0.85 0.8599

0.4

0.6

0.8

Epochs

Accuracy

FeatureExtractor MobileNet-V1

Training

Validation

Epochs

Loss

FeatureExtractor MobileNet-V1

Training

Validation

Figure 5: Accuracy and Loss for Training and Validation:

FeatureExtractor MobileNet-V1.

1. Inference Time: This refers to the inference time

on the edge device, which is the duration taken by

the model from receiving the input to delivering

the evaluated output. It provides insights into the

speed or efﬁciency of the model. As depicted in

Table 4, the inference time of quantized models

was signiﬁcantly reduced for both FP32 and FP16

0.6

0.8

Epochs

Accuracy

FineTuned MobileNet-V1

Training

Validation

Epochs

Loss

FineTuned MobileNet-V1

Training

Validation

Figure 6: Accuracy and Loss for Training and Validation:

FineTuned MobileNet-V1.

quantized models scoring 0.16 and 0.17 seconds

respectively in comparison with the model with-

out quantization which scored an average infer-

ence time of 0.62 seconds. This improvement en-

hances the suitability of these models for deploy-

ment on edge devices.

ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy

0.4

0.6

0.8

Epochs

Accuracy

FeatureExtractor MobileNet-V2

Training

Validation

Epochs

Loss

FeatureExtractor MobileNet-V2

Training

Validation

Figure 7: Accuracy and Loss for Training and Validation:

FeatureExtractor MobileNet-V2.

2. Quantization Effect on Accuracy: Table 4 illus-

trates that the testing accuracy of the MobilNet-

V2 Finetuned model was 86%, while both quan-

tized models with FP16 and FP32 achieved

slightly reduced accuracy at 85% for both. How-

ever, the quantization effect on the chosen model

did not signiﬁcantly impact accuracy.

3. Model Size: This indicates the storage demand

of the model on the edge device, inﬂuencing both

device storage capacity and download bandwidth.

As indicated in Table 4, the FP32 and FP16 mod-

els exhibited smaller sizes compared to the base

model, which was 12.04 megabytes. FP16 had

the smallest size of 3.02 megabytes. This reduc-

tion in size is indicative of the efﬁciency in model

size achieved through quantization.

Taking into account all the evaluation criteria, it was

observed that the test accuracy and inference time for

both FP16 and FP32 models were nearly identical.

However, the model size of the FP16 variant was ap-

proximately half the size of the FP32 model. There-

fore, FP16 models were selected for deployment on

0.6

0.8

Epochs

Accuracy

FineTuned MobileNet-V2

Training

Validation

Epochs

Loss

FineTuned MobileNet-V2

Training

Validation

Figure 8: Accuracy and Loss for Training and Validation:

FineTuned MobileNet-V2.

the Edge Raspberry Pi device, striking a balance be-

tween accuracy and model size efﬁciency.

6.3 Architecture Ability to Enrich

Access Requests

To evaluate the efﬁcacy of the proposed architec-

ture in handling access requests to images based on

their contents, we thouroughly evaluated it using dif-

ferent scenarios targeting various policies.These sce-

narios encompassed requests directed at both pre-

classiﬁed images (where the image contents’ classiﬁ-

cation information was previously saved in the ontol-

ogy), and new, non-classiﬁed images (where the im-

age content information has not been identiﬁed yet).

The proposed Content-Driven access control architec-

ture demonstrated its remarkable capability to pre-

cisely enrich requests with image content properties

across the spectrum of testing scenarios, each aligned

with different stored policies. As a result, our pro-

posed architecture unequivocally proved its effective-

ness in dynamically adapting to diverse access re-

quest scenarios and its agility in incorporating real-

CDAC: Content-Driven Access Control Architecture for Smart Farms

Table 4: On Device (Raspberry Pi) Model Compression Effect on Fine-Tuned MobileNet V2.

Model Model Size (MB) Avg Inference Time (Sec) Test Accuracy

Original Model 12.04 0.62 86%

FP32 5.98 0.16 85%

FP16 3.02 0.17 85%

time classiﬁcation outcomes into access decisions,

thereby showcasing its robust performance in enhanc-

ing smart farm system security.

6.4 Class Speciﬁc Threshold

Adjustment Effect

We evaluated the effect of adjusting different lev-

els of class-speciﬁc thresholds for the Crops major

class when the architecture receives access requests

to images containing crops and targeting policy-2 de-

scribed in Table 2. The test dataset contains a total

of 750 images of different crop types out of 4500 to-

tal test images. Based on the model classiﬁcation re-

sults with different conﬁdence scores and using a cus-

tomizable threshold speciﬁc to Crops, we evaluated

the number of images that were classiﬁed as a type of

crop and are considered suitable for access authoriza-

tion as follows:

• True Positives (TP): Number of images that were

classiﬁed as at least one of the crop classes with

conﬁdence score ≥ crop class speciﬁc threshold

and are indeed crop images.

• False Positives (FP): Number of images that were

classiﬁed as one of the crop classes with conﬁ-

dence score ≥ crop class speciﬁc threshold and

are not crop images.

Table 5 illustrates how different thresholds inﬂu-

ence the access authorization to images based on their

content classiﬁcation conﬁdence scores. As for the

100% threshold, fewer images are deemed suitable

for access authorization due to stricter criteria, lead-

ing to a decrease in access to 14.7% of crop images

and eliminating all access to misclassiﬁed images. On

the other hand, using a smaller threshold of 20% in-

creases the number of crop images eligible for access

to 84.5% but includes more misclassiﬁed images to

31.9%. Notably, Thresholds ranging from 40% to

80% showed a balanced decrease in both the true pos-

itive and false positive access percentage. This under-

scores the critical role of threshold selection in deter-

mining the trade-off between granting access to rele-

vant images and mitigating the risk of granting access

to irrelevant or misclassiﬁed images.

Given this observation, the class-speciﬁc thresh-

old emerges as a crucial element in mitigating in-

correct image classiﬁcation outcomes, particularly

concerning the security sensitivity levels associated

with the content. Consequently, integrating the class-

speciﬁc threshold directly into the policies would en-

hance efﬁciency further. This approach will allow

smart farm administrators to directly formulate poli-

cies that describe the desired conﬁdence levels for im-

age classiﬁcation, aligning with the speciﬁc security

requirements and content sensitivities of the farm im-

ages.

Table 5: Access Authorization Percentage based on Class

Speciﬁc Thresholds.

Threshold True Positives False Positives

% % %

20 84.5 31.9

40 78 23.8

60 73.6 17.5

80 71.1 11.2

100 14.7 0

6.5 Average Request Evaluation Time

The higher Request Evaluation time for pre-classiﬁed

images, as indicated in Table 6, can be attributed

to the time required for Inference from Ontology.

This inference process involves extracting informa-

tion from a structured representation of image con-

tent, and its duration can be inﬂuenced by the size

and complexity of the ontology. The average time for

Inference from Ontology ranges from 1.713 to 1.797,

reﬂecting the intricate nature of ontology-based rea-

soning. The ontology’s size and the depth of seman-

tic relationships within it can contribute to a slightly

longer processing time. On the other hand, the Re-

quest Evaluation time for non-classiﬁed images is no-

tably lower, ranging from 0.56 to 0.67. In this case,

the system relies on the image classiﬁer, which tends

to operate more swiftly. The classiﬁer’s efﬁciency

in inferring image content contributes to the shorter

Request Evaluation time compared to the ontology-

based approach. This nuanced difference in process-

ing times underscores the trade-off between ontology-

based reasoning and the efﬁciency of image classi-

ﬁers, highlighting the system’s adaptability to dif-

ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy

Table 6: Average Total Content-Driven Access Request Evaluation Time of Pre-Classiﬁed & non-Classiﬁed Images.

Classiﬁed Image Non-Classiﬁed Image

Inference(Ontology) Request Evaluation Inference(Classiﬁer) Request Evaluation

Low 1.713 2.113 0.17 0.56

Avg 1.752 2.192 0.18 0.63

High 1.797 2.267 0.20 0.67

ferent scenarios and requirements. Despite the vari-

ations, the overall system maintains practical time

manageability, ensuring effective access control in di-

verse smart farm image scenarios.

6.6 Limitation

One limitation of our proposed approach to plant im-

age classiﬁcation lies in the lack of images containing

occluded or underdeveloped plants in our dataset. It

is important to train the model on diverse conditions

of plants which is common in smart farm images.

For instance, occluded plant images might include in-

stances where leaves or other objects partially obscure

the plant, challenging the model to correctly classify

the image content. Similarly, images of underdevel-

oped plants could feature younger crops with fewer

leaves or less visible growth patterns, necessitating

the model’s ability to differentiate between varying

stages of plant development. The absence of this type

of plant image could adversely affect the classiﬁcation

accuracy of real smart farm images. Consequently,

misclassiﬁcations may occur, potentially increasing

undesired access request decisions.

7 RELATED WORK

Over the years, researchers in access control have

proactively harnessed the capabilities of machine

learning to derive more efﬁcient solutions. Numer-

ous research endeavors have advocated for ML-based

approaches to formulate access control policies that

exhibit greater robustness compared to conventional

methods (Bui and Stoller, 2020), (Cotrini et al., 2018),

(Abu Jabal et al., 2020), (Karimi et al., 2021), (Zhou

et al., 2019). Additionally, rather than employing

ML to address the entire problem, it can be lever-

aged to aid in resolving speciﬁc aspects within the

access control domain. Therefore, Researchers have

utilized ML advancements to automate cumbersome

tasks in access control, including the extraction of at-

tributes from information or text presented in plain

natural language (Alohaly et al., 2018),(Alohaly et al.,

2019b), (Abdi et al., 2022), (Alohaly et al., 2019a),

(Sandhu, 2021), the mapping between roles and per-

missions (Zhou et al., 2019), the extraction of secu-

rity rules from access logs (Cotrini et al., 2018), and

(Karimi et al., 2021), and even the extraction of access

control policies from user stories (Sandhu, 2021).

Furthermore, researchers advocated for the inte-

gration of machine learning into the decision-making

process of access control, as demonstrated in recent

studies such as (Liu et al., 2021) and (Nobi et al.,

2022). These studies underscore the beneﬁts of lever-

aging an ML model for heightened accuracy in access

control decision-making. In this paradigm, access de-

cisions pivot on a trained ML model rather than a pre-

deﬁned access control policy. Typically, these models

render access control decisions—granting or deny-

ing access—by analyzing user and resource metadata

along with associated attributes. These metadata and

attributes encapsulate user and resource features that

the ML model learns for subsequent access decisions.

Nevertheless, as far as our knowledge extends,

there has been a notable absence of the application of

machine learning techniques tailored to facilitate in-

formed decision-making in access control scenarios.

8 CONCLUSION

In this paper, we introduced a content-driven access

control architecture that addresses the critical issue of

regulating access to the extensive collection of smart

farm images. Unlike traditional approaches, this

novel architecture prioritizes regulating access based

on the visual content of the images. The key com-

ponents of this architecture include an edge machine-

learning classiﬁcation model and a semantic attribute-

based access control mechanism. The edge classiﬁca-

tion model operates efﬁciently by providing a quick

and lightweight solution for classifying images near

their source in the smart farm. Its speed and accu-

racy make it a valuable tool for enhancing the overall

performance of the access control system. Moreover,

the semantic attribute-based access control mecha-

nism adds a layer of sophistication by enriching ac-

cess control requests with detailed image content in-

formation. This enhancement reduces the likelihood

CDAC: Content-Driven Access Control Architecture for Smart Farms

of undesired access to smart farm image data, ensur-

ing that only authorized personnel can retrieve spe-

ciﬁc image content.

The effectiveness of this architecture is demon-

strated by its ability to enrich access requests with

necessary image information, facilitating a more in-

formed decision-making process based on the visual

content of the images. Additionally, the edge clas-

siﬁcation model’s lightweight, fast, and accurate na-

ture contributes to the overall efﬁciency of the sys-

tem. Furthermore, the average total time taken to

evaluate access requests for both pre-classiﬁed and

new images is remarkably small, with only slight ad-

ditional time required for pre-classiﬁed images due

to the time taken for ontology inference. This ef-

ﬁciency ensures a seamless and swift access control

process, contributing to the overall success and prac-

ticality of the proposed architecture in a smart farm

environment.

REFERENCES

Abdi, A. I., Eassa, F. E., Jambi, K., Almarhabi, K., Khe-

makhem, M., Basuhail, A., and Yamin, M. (2022).

Hierarchical blockchain-based multi-chaincode ac-

cess control for securing iot systems. Electronics,

11(5):711.

Abu Jabal, A., Bertino, E., Lobo, J., Law, M., Russo, A.,

Calo, S., and Verma, D. (2020). Polisma-a framework

for learning attribute-based access control policies. In

Computer Security–ESORICS 2020: 25th European

Symposium on Research in Computer Security, ES-

ORICS 2020, Guildford, UK, September 14–18, 2020,

Proceedings, Part I 25, pages 523–544. Springer.

Alohaly, M., Takabi, H., and Blanco, E. (2018). A deep

learning approach for extracting attributes of abac

policies. In Proceedings of the 23nd ACM on Sym-

posium on Access Control Models and Technologies,

pages 137–148.

Alohaly, M., Takabi, H., and Blanco, E. (2019a). Auto-

mated extraction of attributes from natural language

attribute-based access control (abac) policies. Cyber-

security, 2(1):2.

Alohaly, M., Takabi, H., and Blanco, E. (2019b). Towards

an automated extraction of abac constraints from nat-

ural language policies. In ICT Systems Security and

Privacy Protection: 34th IFIP TC 11 International

Conference, SEC 2019, Lisbon, Portugal, June 25-27,

2019, Proceedings 34, pages 105–119. Springer.

Bera, S. and Shrivastava, V. K. (2020). Analysis of vari-

ous optimizers on deep convolutional neural network

model in the application of hyperspectral remote sens-

ing image classiﬁcation. International Journal of Re-

mote Sensing, 41(7):2664–2683.

Bui, T. and Stoller, S. D. (2020). Learning attribute-based

and relationship-based access control policies with

unknown values. In International Conference on In-

formation Systems Security, pages 23–44. Springer.

Cheng, J., Wu, J., Leng, C., Wang, Y., and Hu, Q.

(2017). Quantized cnn: A uniﬁed approach to ac-

celerate and compress convolutional networks. IEEE

transactions on neural networks and learning systems,

29(10):4730–4743.

Cotrini, C., Weghorn, T., and Basin, D. (2018). Mining

abac rules from sparse logs. In 2018 IEEE Euro-

pean Symposium on Security and Privacy (EuroS&P),

pages 31–46. IEEE.

Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-

Fei, L. (2009). Imagenet: A large-scale hierarchical

image database. In 2009 IEEE conference on com-

puter vision and pattern recognition, pages 248–255.

Ieee.

Howard, A. G. (2017). Mobilenets: Efﬁcient convolutional

neural networks for mobile vision applications. arXiv

preprint arXiv:1704.04861.

Karimi, L., Aldairi, M., Joshi, J., and Abdelhakim, M.

(2021). An automatic attribute-based access control

policy extraction from access logs. IEEE Transactions

on Dependable and Secure Computing, 19(4):2304–

2317.

Liu, A., Du, X., and Wang, N. (2021). Efﬁcient access

control permission decision engine based on machine

learning. Security and Communication Networks,

2021(1):3970485.

Nobi, M. N., Krishnan, R., Huang, Y., Shakarami, M., and

Sandhu, R. (2022). Toward deep learning based access

control. In Proceedings of the Twelfth ACM Confer-

ence on Data and Application Security and Privacy,

pages 143–154.

Sandhu, R. (2021). Access control policy generation from

user stories using machine learning. In Data and

Applications Security and Privacy XXXV: 35th An-

nual IFIP WG 11.3 Conference, DBSec 2021, Cal-

gary, Canada, July 19–20, 2021, Proceedings, volume

12840, page 171. Springer Nature.

Yassin, G. I. and Ramaswamy, L. (2022). Effective & ef-

ﬁcient access control in smart farms: Opportunities,

challenges & potential approaches. ICISSP, pages

445–452.

Yassin, G. I. and Ramaswamy, L. (2023). Harvesting secu-

rity: A semantically enriched access control architec-

ture for smart farms. In 2023 5th IEEE International

Conference on Trust, Privacy and Security in Intelli-

gent Systems and Applications (TPS-ISA), pages 335–

343. IEEE.

Zhou, L., Su, C., Li, Z., Liu, Z., and Hancke, G. P. (2019).

Automatic ﬁne-grained access control in scada by ma-

chine learning. Future Generation Computer Systems,

93:548–559.

ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy