Abstract:
"
In a growing world of social media where platforms completely rely on images to share
content, image steganography is a convenient candidate for criminals to do secret
communications. So it is important that detecting these images through image steganalysis
so that the illegal activities could be monitored and resolved. In the domain, the
conventional approaches relied on manual feature extraction techniques but with
Convolutional Neural Network (CNN) it was automated making CNN the trending topic.
In this work, approaches that aren’t widely explored in the image steganalysis domain were
analyzed which is Transfer Learning(TL). Also considering the domain, JPEG steganalysis
was scarce so it was scoped into this research. The concept of Transfer Learning using
recent ImageNet models such as EfficientNet, MobileNetV3, DenseNet and MixNet and
attempts on developing ensemble models using the above architectures were experimented.
From these multiple experiments, EfficinetNet showed better results than the rest.
Ensembling these models also showed positive results but a higher computational cost was
required. JPEG Image steganalysis tool was developed where a cybersecurity forensic user
can upload a JPEG image and get a detection for their work. This document will show how
these above decisions and conclusions were made and finally as a major contribution how
Transfer Learning impacts JPEG image steganalysis would be conveyed to the research
community."
