Dipartimento di Informatica

Homographic attack is a well-known squatting technique uses the visual similarity between characters to lure users into phishing campaigns (e.g., word and vvord). In this thesis the student will investigate the use of multi-modal generative models to generate homographic attacks using visual feedback. The visual feedback, obtained with techniques employed in computer vision, will inform the models about the similarity between characters and their context in words to generate words that can increase the effectiveness of the attack. This approach has a huge potential when we consider the internationalized scenarios (URLs), in which words can include a mix of character sets from various alphabets, such as Cyrilic and Persian, as well as the possibility of using multiple font types and sizes when transmitting phishing messages. The student will design and evaluate machine learn models capable of generate such attacks based on the actual graphic similarity between generated candidate words and target words.