Analisis Performa Fitur ATP + CQT pada Algoritma SVM dalam Mendeteksi Serangan Voice Spoofing Replay Attack

Abstract

Automatic Speaker Verification (ASV) is a system that can determine whether a person speaking is the identity he claims to be or not. However, ASV systems are vulnerable to voice spoofing attacks. Voice spoofing attacks can make the ASV system inaccurate in distinguishing between bonafide and spoofed voices. This can be a threat as a counterfeiter can cheat the verification system and perform unwanted actions. One of the voice spoofing techniques is replay attack, which is done by playing a previously recorded speaker's voice. Although fairly simple, replay attacks have proven to be an effective way to deceive ASV systems. For this reason, a system that can detect replay attacks is needed. This study uses Acoustic Ternary Pattern (ATP) and Constant-Q Transform (CQT) feature extraction to recognize sound characteristics so that the difference between bona fide and spoof sounds can be studied. These features are then used to train the Support Vector Machine (SVM) model. Testing was carried out in 3 scenarios, namely: using the ATP feature only, using the CQT feature only, and using a combination of both. The combination of the CQT feature in the ATP feature with the SVM algorithm can improve the model performance in terms of accuracy, EER, and model sensitivity in detecting voice spoofing, especially in replay attacks. The EER value obtained on the ATP+CQT feature with the SVM model is 6.158% with an accuracy of 95.2%.