Abstract:
"
The internet has become a primary source of information for most of its users. While
this may come with its advantages and disadvantages. One such disadvantage is the
existence of false news or Fake news. While manual fact-checkers are used to
combat fake news, they are not as scalable as automated systems. Fake news has
spread far and wide and has long reaching consequences in present day society
worldwide, Images which are included in news stories are eleven times more likely
to be shared compared to ones that are not. Existing research has identified image
manipulation and native emotion within images to be factors of an image being a
fake news image.
The research introduces a more accurate fake news image detection system, while the
current systems have sound accuracy, the tests conducted in ablation studies show
that the image manipulation performs poorly when tried on previous systems. The
research combines deep learning and transfer learning in order to bring about a more
accurate system to previous systems.
This research pursued a novel method of classifying fake news images with the usage
of deep learning and transfer learning, the model utilises is a ResNet101 model,
through this the model has been previously trained on the ImageNet database which
consists of over fourteen million images, this enables the have a significant edge in
terms of accuracy when classifying images. However, that alone is not a determining
factor as the residual blocks in the ResNet achieves an accuracy of 86.9% which
enables it to have higher accuracy to previous systems by 2.3%. "
