Yeniden oynatma saldırıları altında ses doğrulama sisteminin kırılganlığı ve sentetik ses tespitinin incelenmesi

Security issues have become major subject along with developments in the speaker verification system. Previous studies have shown that performance is still not fully proven for vulnerability. Simple and easy methods may be preferred to attack however the verification system can be attacked with high-tech infrastructure. The first example of this method is "Replay Attack". In this thesis, firstly we downloaded data from the Internet in ways that everyone can easily success. Then the data has enrolled into the system that was previously created as universal background model (UBM). Operability of the system have been proved by tests with same people voice data and then system was exposed to attack again with most similar speech that obtained from other speakers. The results also showed that, even it is not high rated but spoofing attack may cause false acceptance by the verification system. After that, noise added to simulate effect of electronic devices and that added noise was found to increase the system performance. With this work it is aimed to point the open field in the anti-spoofing area against the replay attack. Then the specification of synthetic speech system and the differences according to natural sound is investigated. How it is used as spoofing attack and the results are also represented.