| KEYWORD |
Evaluation of smart speakers' vulnerability to sound-squatting using quasi-homophones
keywords CYBERSECURITY, MACHINE LEARNING, ARTIFICIAL NEURAL NETWORKS, PHISHING
Reference persons MARCO MELLIA
External reference persons Idilio Drago
Rodolfo Valentim
Research Groups SmartData@PoliTO, Telecommunication Networks Group
Thesis type EXPERIMENTAL
Description The increase in the usage of voice controls pushes the growth of sound-squatting attacks. Sound-squatting is a phishing technique that tries to trick users by leveraging words with pronunciation perceptively similar to targets. Sound-squatting is, in fact, generic and applies to the case of listeners that have to write (or interpret) a word pronounced by another person, such as, the speech recognition process used in smart speakers. In this thesis the student will evaluate if a data-driven generation of quasi-homophones can systematically trick speech recognition systems that power smart-speakers, specifically Alexa Skill Toolkit. The evaluation process will involve the collection of smart-speakers speech commands and the generation and evaluation of miss-interpretations made by the speech recognition system.
Required skills - Interest in cyber-security
- Interest in machine learning and AI algorithms
- Good programming skill (Python)
- Knowledge of AI, Transformers, DNN
Deadline 23/11/2023
PROPONI LA TUA CANDIDATURA