한성규, 정수환,

Vol. 35, No. 3, pp. 555-562, 6월. 2025
10.13089/JKIISC.2025.35.3.555, Full Text:
Keywords: Audio Deepfake, real-time detection, application, Light-Weight AI Model, voice activity detection
Abstract

With the advancement of AI-based voice synthesis technologies and the growing prevalence of voice cloning services, the threat of audio deepfakes is increasing, highlighting the need for Real-Time Detection solutions. This study implements a mobile application capable of capturing streaming audio from external applications in real time and detecting deepfakes. To ensure stable performance in resource-constrained mobile environments, a pre-trained lightweight AI model is adopted, and post-processing techniques—Voice Activity Detection (VAD) and Moving Average—are combined to improve detection accuracy and real-time responsiveness. The system also evaluates performance across various input durations to identify a configuration that balances responsiveness and reliability. Experimental results show that the proposed application achieved 64% fake detection accuracy and 89% real speech detection accuracy with a 2.1-second input and post-processing, demonstrating its capability to satisfy both real-time and accuracy requirements. These findings confirm the practical feasibility of real-time mobile audio deepfake detection. Future work will focus on enhancing detection stability under diverse environmental conditions and extending the system to other applications such as call security, voice authentication, and video conferencing.

Statistics
Cite this article