RealACaller: Real-Time Arabic Audio Deepfake Detection App for Smartphones

Audio deepfakes present major threats to privacy and security, especially in underrepresented languages such as Arabic. This paper presents a real-time deepfake detection model for Arabic audio, using Facebook's Wav2Vec2 self-supervised learning framework. The model was fine-tuned on the Arabic Audio Deepfake Dataset (ArAD), which includes diverse Arabic dialects including Levantine and Modern Standard Arabic, achieving high accuracy and efficient performance. Unlike existing solutions that primarily focus on the English language, this research provides a crucial and effective solution for over 300 million Arabic speakers. This study discusses the system's design, implementation, and evaluation, focusing on the fact that it can be deployed on resource-constrained devices such as smartphones.