Phonetic-DeepKANet: a robust audio spoofing detection framework for English and Arabic

Area of Science:

• Cybersecurity
• Signal Processing
• Machine Learning

Background:

• Audio spoofing attacks, including deepfakes, threaten automatic speaker verification systems, causing data breaches and financial fraud.
• Existing countermeasures struggle with generalization, particularly against unknown spoofing methods and in underrepresented languages like Arabic due to dataset scarcity.

Purpose of the Study:

• To propose a novel dual-modality approach, Phonetic-DeepKANet (PDK-Net), for reliable audio spoofing detection in both English and Arabic.
• To address the scarcity of Arabic audio spoofing datasets by introducing a new dataset for research advancement.

Main Methods:

• PDK-Net integrates TransRawNet (TR-Net) for deep feature extraction and an acoustic-phonetic feature extraction module.
• Features are concatenated and classified using a Kolmogorov Arnold Network (KAN).
• The method is evaluated on ASVspoof-2019 LA, ASVspoof-2021 LA, ASVspoof-2021 DF, and a newly created Arabic audio spoofing dataset.

Main Results:

• PDK-Net achieved state-of-the-art performance, with the best min-tDCF of 0.09 on ASVspoof-2019 LA and 0.14 on ASVspoof-2021 LA.
• For the Arabic dataset, PDK-Net attained an Equal Error Rate (EER) of 8.06%.
• The method ranked best for LA attacks in ASVspoof-2021 and third-best for DF attacks, demonstrating effectiveness against unknown spoofing types.

Conclusions:

• The proposed PDK-Net offers a robust and generalizable solution for audio spoofing detection across multiple languages.
• The introduction of an Arabic dataset significantly contributes to research on underrepresented languages.
• PDK-Net effectively combats various spoofing attacks, including deepfakes, voice conversion, and text-to-speech synthesis.