Abstract
The latest improvements regarding the online world have come with great benefits, but, as well as dangerous drawbacks (i.e., web-malware). This article proposes to investigate the reliability and accuracy of a novel web-malware detection method by using images and deep learning. The web links are transformed into colored and grayscale images and then a deep learning model is applied to extract relevant features and classify the links. We managed to calibrate and compare a 3-layer convolutional network and the well-known VGG16, ResNet50V2 and ResNet101 models. Moreover, we added a Long-Short Term Memory layer with the focus of improving performance. The approach was tested on two datasets and our best results reached 96.82 % accuracy surpassing the performance of its counterparts on the largest dataset. Moreover, there are not significant differences between the classification done on the grayscale and on the RGB images. However, the Long-Short Term Memory layer may be influenced on how we have generated the input images, in a row-by-row or in a column-by-column manner.
Citare
@Inproceedings{Coste2024MaliciousWL,
author = {C. Coste},
booktitle = {2024 IEEE/WIC International Conference on Web Intelligence and Intelligent Agent Technology (WI-IAT)},
title = {Malicious Web Links Detection Based on Image Processing and Deep Learning Models},
year = {2024}
}