Malware Detection for Visualized Images Using Hybrid Fast R-CNN
and Transformation Models
Swathi Anil, Ananya S Mallia and Manazhy Rashmi
Department of Electronics and Communication, Amrita Vishwa Vidyapeetham, Amritapuri, Kollam, India
Keywords: Malware Detection, Visualized RGB Images, Fast R-CNN, Transformation Model, Hybrid Approach, Feature
Extraction, Ensemble Learning
Abstract: Malware visualization techniques are becoming increasingly sophisticated, posing significant challenges to
traditional detection systems. To address this, we propose a novel Hybrid Fast R-CNN and Transformation
Model (HFRTM) framework for the accurate detection and classification of malware in visualized RGB
images and malicious network behaviours. The HFRTM integrates Fast Region-based Convolutional Neural
Networks (Fast R-CNN) for efficient malware pattern detection and localization with transformation models
to enhance feature extraction by capturing complex variations in malware appearances. Key enhancements
like fine-tuning of transformation models on a specialized target dataset, leveraging pre-trained weights to
accelerate convergence and mitigate overfitting. This ensemble architecture demonstrates superior accuracy
and robustness, effectively distinguishing malware from benign data even in challenging scenarios. To
validate the efficacy of HFRTM, extensive experiments were conducted on a benchmark malware
visualization dataset. The proposed method achieved a detection accuracy of 98.7%, significantly
outperforming existing state-of-the-art methods. The results highlight the practical applicability of HFRTM
in real-world cybersecurity scenarios, offering an advanced and reliable solution for combating sophisticated
malware threats.
1 INTRODUCTION
Malware detection and classification have become
increasingly vital in the realm of cybersecurity as
cyber threats continue to evolve in complexity and
sophistication. Traditional malware detection
systems, which rely heavily on signature-based
methods, are proving inadequate in the face of novel
and more sophisticated attacks. These challenges
necessitate the development of more advanced
techniques that can adapt to the changing landscape
of malware. One such promising approach is the use
of visualized malware images, where malware
binaries are transformed into RGB images to expose
patterns and anomalies that are difficult to detect with
conventional methods. These visual patterns are
effectively analysed using deep learning models
specifically designed for image analysis, enhancing
detection accuracy and robustness against diverse and
emerging threats.
Visualized malware images introduce a
groundbreaking approach to analysing and
classifying malicious software. Unlike traditional
code-based analysis methods, this technique
transforms malware binaries into RGB images,
unveiling intricate patterns and textures that signal
malicious activity. These subtle visual cues, often
invisible to conventional methods, can be effectively
captured and analysed using advanced deep learning
architectures. The Fast Region-based Convolutional
Neural Network (Fast R-CNN) has demonstrated
remarkable capability in detecting and localizing
malware patterns within visualized images and the
associated feature-extracted data. Its efficiency and
accuracy make it a powerful tool in the fight against
evolving cyber threats. However, a significant
challenge lies in ensuring these models can generalize
effectively across diverse malware types. This is
where the integration of transformation models
becomes indispensable. By enhancing feature
extraction, transformation models enable the
detection framework to adapt to the wide variety of
malware appearances, ensuring robust and reliable
classification across complex datasets.
This study proposes a Hybrid Fast R-CNN and
Transformation Model (HFRTM) for enhanced
574
Anil, S., S Mallia, A. and Rashmi, M.
Malware Detection for Visualized Images Using Hybrid Fast R-CNN and Transformation Models.
DOI: 10.5220/0013582000004664
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 1, pages 574-581
ISBN: 978-989-758-763-4
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

malware detection. By integrating transformation
models, HFRTM improves feature extraction,
capturing complex malware variations that simpler
models miss. Fine-tuning on a target dataset with pre-
trained weights accelerates training and reduces
overfitting, ensuring better generalization. The
ensemble architecture enhances detection accuracy,
minimizes false positives, and maintains a high
detection rate. Validated on benchmark datasets,
HFRTM outperforms existing methods, proving its
effectiveness as a robust defence against evolving
malware threats.
2 RELATED WORK
The evolution of deep learning and feature selection
techniques has significantly transformed the
landscape of malware detection, creating a narrative
of continuous innovation and refinement. At the
forefront, Atacak (Atacak, et al. , 2019) laid the
groundwork by introducing the FL-BDE system, a
fuzzy logic-based dynamic ensemble for Android
malware detection. This pioneering system integrated
six machine learning techniques, such as decision
forests and neural networks, to enhance classification
accuracy, setting the stage for multi-technique
ensemble approaches.
Building upon this foundation, Masum et al.
(Masum, Faruk, et al. , 2022) tackled the specific
challenge of ransomware detection. By employing
feature selection alongside machine learning methods
like Random Forest and neural networks, their work
provided robust threat categorization, demonstrating
the potential of tailored methodologies in combating
specialized malware threats. Inspired by the
effectiveness of feature selection, Alomari et al.
(Alomari, Nuiaa, et al. , 2023)extended this approach
to high-dimensional malware data, developing a
sophisticated system that combined LSTM-based
deep learning models with correlation-based feature
selection. This solution addressed the growing
complexity and volume of malware datasets.
In parallel, Kumar (Kumar, 2023) explored
innovative architectures like CNN-BiLSTM to
counteract the increasing sophistication of modern
malware. Kumar’s work underscored the importance
of well-curated datasets, reinforcing the need for
robust data preparation in achieving high detection
performance. While these advancements focused on
text-based feature extraction, a paradigm shift
occurred with the introduction of image-based
malware detection.
Nataraj et al. (Nataraj, Karthikeyan, et al. , 2011),
(Nataraj, and, Manjunath, 2016) revolutionized the
field by representing malware binaries as images,
uncovering family-specific patterns and initiating a
new line of research in visualized malware analysis.
Building on their pioneering efforts, Han et al. (Han,
Kang, et al. , 2014), (Han, Lim, et al. , 2015)
incorporated image similarities and entropy maps to
achieve more precise classification, demonstrating
the versatility of image-based approaches.
Expanding on the concept of visual analysis, Liu et al.
(Liu, Wang, et al. , 2017) employed clustering
techniques with grayscale images, offering a novel
perspective on effective classification. Fu et al. (Fu,
Xue, et al. , 2018) took this a step further by
highlighting unique malware features through colour
image analysis, adding depth to the visualization
approach. The journey of image-based malware
detection continued with Singh et al. (Singh, Handa,
et al. , 2019), who applied CNN models to visualize
malware, demonstrating the power of deep learning
in extracting meaningful patterns from images.
The exploration of image classification
techniques branched out into other domains, with
region-based approaches like the watershed
transform playing a pivotal role. A watershed-based
segmentation method (Pawar, Perianayagam, et al. ,
2017) highlighted its ability to delineate regions of
interest, improving classifier performance in
challenging environments. The successes in static
image classification influenced dynamic tasks, such
as sign language recognition. Here, hybrid models
combining CNNs and Recurrent Neural Networks
(RNNs) (Renjith, Manazhy, et al. , 2024)
demonstrated their efficacy by capturing both spatial
and temporal features, achieving higher accuracy in
classifying dynamic gestures.
Advancements in hybrid learning models have
also contributed significantly to interpretability.
Harishankar et al. (Harishankar, Anoop, et al. , 2024)
introduced an explainable hybrid model for Indian
food image classification, combining feature
extraction with explainable AI techniques. This
approach enhanced both understanding and
reliability, echoing the need for transparency in
classification decisions. Extending these principles to
malware detection, Harishankar et al. (Harishankar,
Anoop, et al. , 2024) presented an ensemble-based
approach for classifying and interpreting dynamic
malware behaviours, achieving improved accuracy
and reliability.
Malware Detection for Visualized Images Using Hybrid Fast R-CNN and Transformation Models
575

3 OVERVIEW
3.1 Architecture
The architecture of the proposed Hybrid Fast R-CNN
and Transformation Model (HFRTM) is specifically
designed to maximize malware detection accuracy in
visualized RGB images by effectively combining
Fast R-CNN and Discrete Wavelet Transform (DWT)
models. The Fast R-CNN component first generates
region proposals and localizes potential malware
patterns within the images. Using a convolutional
neural network (CNN), it extracts spatial features that
are essential for identifying malware signatures.
These features are critical for reducing false negatives
by accurately pinpointing areas indicative of
malicious activity.
To further enhance detection, the architecture
integrates DWT as a transformation model. DWT
captures intricate variations and subtle anomalies in
malware appearances, enabling the detection of
complex patterns that traditional models might
overlook. This hybrid approach effectively bridges
the gap between localized pattern recognition (Fast R-
CNN) and nuanced anomaly detection (DWT),
providing a comprehensive analysis of malware
behaviour. The model leverages pre-trained weights
and fine-tunes them on a benchmark dataset,
significantly accelerating convergence and reducing
overfitting. This ensures the architecture adapts well
to the dataset's diverse malware classes while
maintaining high precision. After feature extraction,
fully connected layers perform the final classification,
distinguishing benign files from malware with
exceptional accuracy.
3.2 Algorithm and Implementation
The implementation begins with dataset preparation,
where malware binaries are visualized as RGB
images and paired with labeled datasets containing
both benign and malicious samples. This dataset
should represent diverse malware categories such as
Trojans, ransomware, and spyware to ensure the
algorithm's robustness. Preprocessing steps include
normalizing pixel values and resizing the images to a
consistent resolution (e.g., 224x224) to meet the input
requirements of the Fast R-CNN and CNN
components.
The Figure 1 illustrates the hybrid deep learning
model that integrates an autoencoder and Fast R-CNN
for classification. The implementation of the Hybrid
Fast R-CNN and Discrete Wavelet Transform
(HFRTM) algorithm for malware detection begins
with preprocessing visualized RGB images of
malware. The images are normalized and resized to
ensure uniformity across the dataset, enabling
consistent input for the algorithm. In the initial stage,
the Fast R-CNN component is employed to detect and
localize potential malware patterns. Fast R-CNN
generates region proposals from the input images,
which are processed through a convolutional neural
network (CNN) to extract spatial features, such as
texture and intensity variations. These features are
critical for highlighting regions that may contain
malicious code, effectively narrowing down the areas
of interest.
To enhance the detection capabilities further, the
Discrete Wavelet Transform (DWT) is integrated as
a transformation layer. The DWT decomposes the
feature maps generated by Fast R-CNN into various
frequency components, capturing both global and
local variations in the malware’s visual patterns. This
multi-resolution analysis enables the system to
identify intricate details and subtle anomalies that
might be missed by traditional CNN-based
approaches. The transformed features are refined
through additional convolutional layers, enriching
their representation and making them more suitable
for classification.
The HFRTM algorithm integrates transfer
learning and ensemble learning to enhance training
efficiency, accuracy, and robustness. Pre-trained
weights from large-scale datasets like ImageNet are
fine-tuned on the malware dataset, accelerating
convergence, reducing overfitting, and improving
adaptation to malware-specific features. Extracted
features are processed through fully connected layers
for final classification into benign or malicious
categories. To further strengthen reliability, multiple
HFRTM models are trained on different data splits,
and their outputs are aggregated using majority voting
or weighted averaging. This ensemble approach
mitigates individual model biases, reducing false
positives and false negatives for more robust malware
detection.
The algorithm’s performance has been rigorously
validated on benchmark datasets, achieving
exceptional results. It recorded an accuracy of 98.7%,
precision of 98.5%, recall of 97.8%, and an F1 score
of 98.1%, underscoring its effectiveness in
differentiating between benign and malicious
samples. The HFRTM algorithm’s ability to integrate
region-based detection through Fast R-CNN,
enhanced feature extraction via DWT, and the
robustness of ensemble learning makes it a powerful
tool for tackling the evolving challenges of malware
detection in cybersecurity.
INCOFT 2025 - International Conference on Futuristic Technology
576

Figure 1: Malware detection framework
3.3 Methodology
3.3.1 Data Collection and Image
Preprocessing
The VirusShare dataset, comprising approximately
70,000 RGB images across 25 distinct malware
families at a resolution of 224x224, plays a crucial
role in training and evaluating malware detection
models. These images, derived from binary malware
files, encapsulate structural features that enable
models to learn complex patterns for accurate
classification. The balanced distribution of malware
families ensures that models trained on this dataset
generalize well to a broad spectrum of malware types.
To further improve detection, an autoencoder-based
anomaly detection approach is utilized. By training
on both malicious and benign images, the
autoencoder compresses input data into a lower-
dimensional latent space, effectively capturing
normal patterns while identifying deviations. During
testing, high reconstruction errors highlight
anomalies, often signaling novel or obfuscated
malware.
Figure 2:- Samples of malware image from VirusShare dataset
3.3.2 Autoencoder Training and Outlier
Detection
The proposed approach leverages an autoencoder-
based framework to detect anomalies indicative of
malware within pre-processed RGB images. An
autoencoder, a type of unsupervised neural network,
is trained on the dataset to compress input images into
a compact, lower-dimensional latent representation
and subsequently reconstruct them. The system's
efficacy lies in its ability to measure the
reconstruction error—the difference between the
original input and its reconstruction. During
inference, samples with reconstruction errors
exceeding a predefined threshold are flagged as
potential malware. Higher reconstruction errors
signify deviations from the learned patterns of benign
samples, effectively identifying anomalies that
include novel or obfuscated malware.
To enhance the practicality and robustness of the
method, the autoencoder training process involves
carefully pre-processed image data. Images are
normalized and resized to ensure uniform input, and
the autoencoder is optimized using a loss function
that minimizes reconstruction errors. The threshold
for anomaly detection is dynamically adjusted based
on the distribution of reconstruction errors observed
during validation, ensuring the model adapts to subtle
variations in the dataset while maintaining high
sensitivity to outliers.
4 EXPERIMENTS AND RESULTS
4.1 Test Bench
The experiments were conducted on a high-
performance computing setup optimized for deep
learning tasks, designed to efficiently process large
datasets and train complex models. The experimental
environment included Python-based libraries such as
TensorFlow, PyTorch, and Keras for implementing
the autoencoder and CNN models, while OpenCV
and Scikit-learn were used for image preprocessing
and feature extraction. The VirusShare dataset,
consisting of visualized malware images, was
preprocessed by normalizing and resizing the images
to a uniform resolution, ensuring consistency across
the dataset. A batch size of 32 was used throughout
the training process to balance memory usage and
computational efficiency. The models were
optimized using the Adam optimizer with a learning
rate of 0.0001, ensuring efficient convergence.
Hyperparameter tuning was performed through cross-
validation and grid search techniques to identify the
optimal settings and maximize model performance.
This comprehensive experimental setup allowed for
the evaluation of key performance metrics, including
accuracy, precision, recall, and F1-score, across
different types of malware
Malware Detection for Visualized Images Using Hybrid Fast R-CNN and Transformation Models
577

4.1.1 Performance of Deep Learning Models
The performance of the Fast R-CNN model on the
VirusShare dataset is summarized in Figure 3,
highlighting its effectiveness in malware detection.
The model achieved a strong accuracy rate of 98.2%,
with a corresponding precision of 98.2%,
underscoring its reliability. Upon removing
anomalies from the dataset, accuracy improved to
98.7%, as shown in Figure 6, emphasizing the
positive impact of outlier removal on classification
performance. The recall rate of 97.3% further
confirms the model's robustness in detecting malware
across a variety of malware types, demonstrating its
reliability in real-world scenarios.
Despite the inherent complexity of the VirusShare
dataset, which contains diverse malware types, Fast
R-CNN maintained consistently high performance.
The precision rate remained stable at 98.2%, while
the recall rate of 97.3% reflected the model's ability
to identify malware samples accurately while
minimizing false positives. These results highlight
the model’s effectiveness in handling large, complex
datasets and its reliability in detecting malware
without overwhelming the system with false alerts.
Figure 3 also presents a histogram showing
Precision, Accuracy, F1 Score, and Recall across
multiple malware categories, including Trojan,
Ransomware, Worm, Backdoor, Spyware, and
Adware. These metrics consistently scored above
85%, with precision and accuracy often exceeding
90%, indicating the model’s strong performance in
detecting malware. The consistently high precision
(blue bars) suggests the model is highly effective at
minimizing false positives, ensuring that benign files
are rarely flagged as malicious. Accuracy (light green
bars) remains robust across all malware categories,
reflecting the model’s overall effectiveness in
correctly classifying both malware and benign files.
Additionally, the F1 Score (orange bars) shows a
balanced performance, maintaining high values
above 90% for malware types like Trojan, Virus,
Phishing, and Downloader, which is crucial for
minimizing both false positives and false negatives.
However, slight variations in recall (red bars)
were observed, particularly for malware types such as
Worm, Keylogger, and Exploit, where recall dipped
slightly below the other categories. This indicates that
while the model is strong overall, there is room for
improvement in detecting certain malware types more
reliably. Addressing this recall variation would
further enhance the model’s ability to detect all
malware types with high precision.
Figure 3 :
Results of VirusShare dataset
4.1.2 Performance Benchmarking
The Hybrid Fast R-CNN and Transformation Model
(HFRTM) demonstrates superior performance over
traditional and contemporary models, as evident from
the key performance metrics summarized in Table 1.
Achieving an accuracy of 98.7%, HFRTM
outperforms both Fast R-CNN (98.2%) and the DWT
+ Fast R-CNN model (98.5%). The model excels in
precision, with a value of 98.5%, showcasing its
ability to accurately identify relevant data while
minimizing false positives. Additionally, HFRTM
exhibits a recall of 97.8%, slightly surpassing the
other models, highlighting its enhanced sensitivity in
detecting all relevant instances. With an F1-Score of
98.1%, the HFRTM strikes an optimal balance
between precision and recall, cementing its
superiority in malware detection.
Table 1 :- Performance Benchmarking
MODE
L
ACC
(%)
PREC
(%)
REC
(%)
F1
(%)
Fast R-
CNN
98.2 98.2 97.3 97.7
DWT+
Fast-
RCNN
98.5 98.3 97.6 97.9
HFRT
M
(Propos
ed)
98.7 98.5 97.8 98.1
INCOFT 2025 - International Conference on Futuristic Technology
578

4.2 Test Cases
4.2.1 Test Case 1: Malware Detection
Accuracy
In the first test case, the performance of the Fast R-
CNN model was thoroughly evaluated for detecting
malware and benign files within a comprehensive
dataset. This evaluation focused on the model’s
ability to accurately classify malware threats and
benign instances. The dataset was divided into
training and testing sets, containing a diverse range of
malware types alongside benign files, ensuring robust
testing conditions. The Fast R-CNN model was
trained on these samples, and key performance
metrics such as accuracy, precision, recall, and F1-
score were used to assess its effectiveness.
The model achieved an impressive accuracy rate
of 98.7%, demonstrating its ability to accurately
classify both malware and benign files across the
dataset. Precision reached 98.5%, indicating that the
model maintained a low false positive rate, with
minimal benign files being incorrectly flagged as
malicious. Furthermore, the recall rate of 97.8%
underscored the model’s robustness in identifying a
significant proportion of actual malware threats,
ensuring that most real threats were detected. These
results highlight the model's effectiveness in practical
scenarios, where achieving high accuracy and
precision is crucial for reliable malware detection.
4.2.2 Test Case 2: False Positive Reduction
In the second test case, the Fast R-CNN model was
evaluated specifically for its ability to minimize false
positives, a key factor in improving system reliability
and user experience. False positives can lead to
unnecessary alarms, system slowdowns, and the
wrongful quarantine of benign files, potentially
disrupting daily operations. The model was tested on
benign files with slight variations to assess how well
it handled legitimate yet unconventional data. The
results showed that the model achieved a precision
rate of 98.5%, significantly reducing the false positive
rate and ensuring that benign files were correctly
classified.
This high precision is especially critical in large-
scale enterprise environments, where numerous
legitimate operations occur continuously. By
minimizing false positives, the model enhanced
operational efficiency, allowing users to trust the
system’s results without the need for frequent manual
intervention or overrides.
Overall, the second test case underscores the
practical value of the Fast R-CNN model in reducing
false positives, allowing it to provide accurate
malware detection while minimizing disruptions to
legitimate operations. This emphasizes the model’s
potential for widespread deployment in environments
that demand high reliability and minimal downtime.
4.3 Evaluation
The evaluation metrics utilized in this study are
consistent with those used in the majority of prior
research. These metrics include the F1-score and
prediction accuracy across various input parameter
settings. The precision of categorization is typically
employed to assess the performance of deep learning
models. Additionally, confusion matrices were used
to compare rates of successful and unsuccessful
predictions. In the confusion matrix, true positives
(TP), true negatives (TN), false positives (FP), and
false negatives (FN) are represented. The primary
metric employed to evaluate the classification
techniques in this study is the F1-score, which
measures the proportion of accurate predictions
across all samples. Accuracy, also referred to as the
true positive rate (TPR), is determined by the ratio of
actual positive outcomes to those predicted by the
classifier, calculated as TP/(TP + FP). Recall, another
crucial metric, is computed as TP/(TP + FN), where
TP indicates the number of true positive predictions
and FN represents the number of relevant samples not
correctly identified. Precision is computed as TP/(TP
+ FP), where TP represents the number of true
positive predictions and FP denotes the number of
false positive predictions. Precision measures the
proportion of correctly identified positive instances
out of all instances predicted as positive by the model.
4.4 Results
The results of the malware detection experiment
using the Fast R-CNN model were highly promising,
with the model achieving an accuracy rate of 98.7%.
This indicates that the model is highly effective at
correctly classifying both malware and benign files.
With a precision rate of 98.5%, the model
significantly reduced the occurrence of false
positives, ensuring that legitimate files were rarely
misclassified as malware. Additionally, the recall rate
of 97.8% demonstrates the model’s strong ability to
identify a large proportion of actual malware
instances, ensuring that threats are not overlooked.
Malware Detection for Visualized Images Using Hybrid Fast R-CNN and Transformation Models
579

These performance metrics confirm the model’s
reliability and efficiency in malware detection,
balancing precision and recall to minimize both false
positives and false negatives. This balance is
particularly important for real-world applications,
where maintaining system security requires accurate
classification without compromising on operational
efficiency. The Fast R-CNN model has proven to be
a robust solution for malware identification,
providing accurate, reliable, and scalable results,
essential for practical deployment in cybersecurity
environments.
5 CONCLUSION AND FUTURE
WORKS
In this study, the HFRTM framework demonstrated
its potential in malware detection, achieving
remarkable performance metrics, including an
accuracy of 98.7%, precision of 98.5%, and recall of
97.8%. These results confirm the framework's
robustness in classifying both malware and benign
files, showcasing its reliability in minimizing false
positives and ensuring comprehensive threat
detection. The balance between precision and recall
highlights the model's suitability for real-world
applications where both false positives and false
negatives must be minimized to maintain system
security.
To enhance the robustness and applicability of the
HFRTM framework, several key improvements are
proposed. Real-time implementation is a primary
focus, enabling continuous malware threat detection
in large-scale environments such as enterprise
networks and cloud platforms. Additionally,
developing lightweight versions for resource-
constrained devices like IoT sensors and mobile
platforms will broaden the framework’s usability.
Integrating transformer-based models will improve
the framework’s ability to capture complex
relationships in malware visualizations, helping to
detect subtle, evolving patterns often missed by
traditional CNN models. Finally, extending the
HFRTM framework to cross-domain applications,
such as medical image analysis for tumour detection
or fraud detection in transaction patterns, will
enhance its versatility and demonstrate its broader
applicability in both cybersecurity and other critical
fields.
REFERENCES
I. Atacak, “An ensemble approach based on fuzzy logic
using machine learning classifiers for android malware
detection,” Applied Sciences, vol. 13, no. 3, p. 1484,
2023.
M. Masum, M. J. H. Faruk, H. Shahriar, K. Qian, D. Lo,
and M. I. Adnan, “Ransomware classification and
detection with machine learning algorithms,” in 2022
IEEE 12th Annual Computing and Communication
Workshop and Conf. (CCWC). IEEE, 2022, pp. 0316–
0322.
E. S. Alomari, R. R. Nuiaa, Z. A. A. Alyasseri, H. J.
Mohammed, N. S. Sani, M. I. Esa, and B. A. Musawi,
“Malware detection using deep learning and
correlation-based feature selection,” Symmetry, vol.
15, no. 1, p. 123, 2023.
M. Kumar, “Scalable malware detection system using
distributed deep learning,” Cybernetics and Systems,
vol. 54, no. 5, pp. 619–647, 2023.
L. Nataraj, S. Karthikeyan, G. Jacob, and B. S. Manjunath,
“Malware images: visualization and automatic
classification,” in Proceedings of the 8th international
symposium on visualization for cyber security, 2011,
pp. 1–7.
L. Nataraj and B. Manjunath, “Spam: Signal processing to
analyze malware [applications corner],” IEEE Signal
Processing Magazine, vol. 33, no. 2, pp. 105–117,
2016.
K. Han, B. Kang, and E. G. Im, “Malware analysis using
visualized image matrices,” The Scientific World
Journal, vol. 2014, 2014.
K. S. Han, J. H. Lim, B. Kang, and E. G. Im, “Malware
analysis using visualized images and entropy graphs,”
International Journal of Information Security, vol. 14,
no. 1, pp. 1–14, 2015.
L. Liu, B.-s. Wang, B. Yu, and Q.-x. Zhong, “Automatic
malware classification and new malware detection
using machine learning,” Frontiers of Information
Technology & Electronic Engineering, vol. 18, no. 9,
pp. 1336–1347, 2017.
J. Fu, J. Xue, Y. Wang, Z. Liu, and C. Shan, “Malware
visualization for fine-grained classification,” IEEE
Access, vol. 6, pp. 14 510–14 523, 2018.
A. Singh, A. Handa, N. Kumar, and S. K. Shukla, “Malware
classification using image representation,” in
International Symposium on Cyber Security
Cryptography and Machine Learning. Springer, 2019,
pp. 75–92.
M. S. Pawar, L. Perianayagam and N. S. Rani, "Region
based image classification using watershed transform
techniques," 2017 International Conference on
Intelligent Computing and Control (I2C2), Coimbatore,
India, 2017, pp. 1-5, doi: 10.1109/I2C2.2017.8321839.
Renjith, S., Manazhy, R., Suresh, M.S.S. (2024).
Recognition of Sign Language Using Hybrid CNN–
RNN Model. In: Hassanien, A.E., Anand, S., Jaiswal,
A., Kumar, P. (eds) Innovative Computing and
Communications. ICICC 2024. Lecture Notes in
INCOFT 2025 - International Conference on Futuristic Technology
580

Networks and Systems, vol 1021. Springer, Singapore.
https://doi.org/10.1007/978-981-97-3591-4_2
C. Harishankar, N. S. Anoop, K. S. Niranjana, A. Biju, V.
Venugopal and S. Vekkot, "An Explainable Hybrid
Learning Model for Indian Food Image Classification,"
2024 15th International Conference on Computing
Communication and Networking Technologies
(ICCCNT), Kamand, India, 2024, pp. 1-6, doi:
10.1109/ICCCNT61001.2024.10724730.
N. Prathapaneni et al., "Dynamic Behaviour analysis and
interpretation of Malware in Android devices using
Ensemble Machine Learning," 2024 3rd International
Conference on Artificial Intelligence For Internet of
Things (AIIoT), Vellore, India, 2024, pp. 1-6, doi:
10.1109/AIIoT58432.2024.10574581.
T. Haritha and Rajesh Kannan Megalingam, "Multiple-
Instance Learning Support Vector Machine Algorithm
based Pedestrian Detection," in Proceedings of the 2020
International Conference, IEEE.
Megalingam, Rajesh Kannan, Gowtham G. Menon, K.
Karthik, and B. Swathi. "Face Detection and Adaptive
Zooming for Intelligent Lecturer Tracking System." In
2023 3rd International Conference on Pervasive
Computing and Social Networking (ICPCSN), pp. 354-
358. IEEE, 2023.
Malware Detection for Visualized Images Using Hybrid Fast R-CNN and Transformation Models
581