Abstract:
The goal of this research is to improve malware classification in the rapidly evolving
field of cybersecurity by applying state-of-the-art deep learning models, namely
ResNet-50 v2, InceptionV3, VGG16, and DenseNet-121. Taking advantage of these
architectures' strong points, this work aims to increase the efficiency and accuracy of
identifying malicious software using a dataset consisting of nine different malware types.
Each model's unique features are carefully examined, and their unique contributions to
classification accuracy within the intricate malware taxonomy are examined. Through a
thorough analysis, the research aims to shed light on the subtle nuances of malware
behavior and features, equipping cybersecurity professionals with advanced tools for
threat identification and mitigation. The results have practical implications for the
creation of more resilient and adaptable security measures in the ongoing fight against
developing cyber threats, in addition to contributing to the scholarly discourse on
malware classification. Among other CNN models used in this study ResNet-50 v2
scored the best accuracy of 86.9%. After that VGG16 and DenseNet-121 showed
promising results with 80.4% and 82.5% accuracy. Traditional approaches like Multilayer
Perceptron, Random Forest, Long Short-Term Memory, K-Nearest Neighbor were also
added to compare between these and Convolutional Neural Network models to find better
solutions to the malware problem in our daily life. Random Forest scored 79% accuracy
which being the highest accuracy among the traditional approaches. Surprisingly
Multilayer Perceptron achieved 94.5% model training accuracy but failed to perform
accordingly while testing the model and scored only 76%. Convolutional Neural Network
outperforming every other traditional approach was an achievement of this study and
Convolutional Neural Network proved to be the better solution than other approaches in
the case of image classification of malware dataset.
