\section{Conclusion}
To protect the numerous users from potential loss for being victim to phishing attack or ID
impersonation and also enable profanity detection tools prevent Unicode based attacks, we have
developed a deep learning based model for discovering homoglyphs as a pre-defense mechanism. Comprehensive analysis of hypertunable parameters of CNN such as stride, drop-out, etc. led to the design of an optimized model. Later, knowledge transfer to our task domain using transfer learning based approach helped us build a superior model. This model achieved a perfect 100\% recall rate and an outstanding 86.5\% precision rate which demonstrates its capability as a homoglyph detector model. We plan to release the model for the security community in near future.