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

Abstract Audio spoofing attacks, specifically deepfakes, are massively used these days to compromise the security of automatic speaker verification-based systems, leading to data breaches and financial scams. Existing audio spoofing countermeasures are not well-generalized and experience issues when detecting unknown spoofing attacks, including deepfake. Moreover, Arabic audio spoofing detection has been largely neglected, primarily due to the scarcity of Arabic language spoofing datasets. This paper proposes a novel dual-modality approach, Phonetic-DeepKANet (PDK-Net), capable of reliable detection of audio spoofing attacks in English and Arabic. The proposed PDK-Net is comprised of a deep feature extraction module incorporating TransRawNet (TR-Net), an acoustic–phonetic feature extraction module, and a Kolmogorov Arnold Network (KAN) classifier. The deep features from TR-Net are complemented with multi-view acoustic–phonetic representations through concatenation and then classified using KAN. This paper also introduces an Arabic audio spoofing dataset to address the limited availability of such datasets and advance the research in audio spoofing detection for underrepresented languages. The proposed method is evaluated utilizing ASVspoof-2019 LA, 2021 LA, and DF, partial spoof, and our Arabic audio spoofing dataset created in this work. Extensive experimentation on multiple datasets, including voice conversion and text-to-speech synthesized samples, algorithm-wise and cross-corpora evaluation demonstrates the effectiveness and generalizability of our method. We attained the best min-tDCF of 0.09 and 0.14 on the ASVspoof-2019 LA and ASVspoof-2021 LA datasets, respectively, compared to baseline models. However, for the Arabic spoofing dataset, the PDK-Net achieved an EER of 8.06%. It is noteworthy that our method performed best for detecting LA attacks over all 41 methods reported in the ASVspoof-2021 challenge. Further, our method registered the third-best EER of 17.55% amongst 33 challenge participants on the ASVspoof-2021 DF set. These results demonstrate the effectiveness and improved generalization of our approach while detecting unknown spoofing attacks, including codec compressions, channel variations, and encoding artifacts.