Abstract:
Character assassination and online harassment have been a major issue in social media
platforms. Most of the time, the target of these attacks are celebrities and politicians.
Further, there have been reports of incidents where memes of nude or intimate photos of
the general public which has been released without the consent of the owner have been
shared in social media platforms. Though there are options to report this kind of incident
in social media platforms itself and organizations like CERT are available for support in
these cases, there is no proper way to search for these images in social media sites.
A research was done in a mixed approach which includes quantitative and qualitative
approaches to come up with a solution to the problem mentioned above. The primary two
purposes of the research were to identify a method to detect images which are potential
character assassinations and to identify the most appropriate technical methodology for the
system. After the analysis, it was decided to use a combination of machine learning, image
processing and text processing techniques, with the use of convolutional, recurrent and
LBHS algorithms. Accordingly, a system was implemented to address the mentioned
problem which contained a face recognition module, a face matching module, image
classification model, two text classification models, an optical character recognition
module and an image scraping module. After the implementation was completed, the built
image model and text models were trained using training datasets to obtain the highest
accuracy possible. Furthermore, a Sinhala language-based hate speech dataset with over
6000 records was created and published publicly as a contribution of this project.
As the ultimate result of the research project, which was conducted throughout nearly a
year, a system to detect image-based character assassination on social media was developed
using machine learning, image processing and text processing techniques. The system can
find the images in the social media platforms which are classified as possible character
attacks, which contains the given face of the end user
