2026-01-07

To tackle the growing spread of deceptive “deepfake” content online, a research team from the Department of Computer Science and Information Engineering at National University of Kaohsiung (NUK) has developed a practical mobile-based Deepfake detection system. Led by Professor Pan Hsin-Tai and student team members Chen Guan-Lin, Chang Pu-Hua, Huang Yi-Jhen, and Liao Guan-Cheng, the system emphasizes portability and real-time usability. It can simultaneously analyze visual and audio consistency on mobile devices, identifying subtle anomalies in face-swapping and voice-altered videos—reducing reliance on human judgment alone.

Professor Pan noted that with the rapid advancement of generative AI technologies, deepfake content has proliferated across social media and messaging platforms, making it increasingly difficult to distinguish real from fake. Such technologies have also been misused for fraud, misinformation, and public opinion manipulation. He emphasized that while deepfake technology itself is not inherently negative, its impact depends on how it is applied. Most existing detection systems rely on large-scale models and high-performance computing, making them inaccessible for everyday users. Therefore, the project was designed to address real-world needs by transforming AI from a lab-based capability into a practical tool for general use.

Team leader Chen Guan-Lin explained that the system is built on a lightweight deep learning model, incorporating multimodal learning techniques to analyze both image and audio data simultaneously. Instead of relying solely on large datasets of authentic images, the team focused on identifying common artifacts found in deepfake content, such as unnatural visual boundaries and inconsistencies. This approach enhances detection performance in real-world scenarios. The system achieves over 90% accuracy on benchmark datasets and has been optimized through model quantization and preprocessing techniques, enabling real-time inference directly on mobile devices. This eliminates the need for cloud uploads, reducing latency while preserving user privacy and ensuring stable performance across different devices and environments.

The project has received multiple external recognitions. It won First Prize in the Information and Electrical Engineering category at the “2025 National AI Project Innovation Competition,” First Prize in NUK’s Departmental Capstone Project Competition, and an Honorable Mention in the Cybersecurity Technology category at the Ministry of Education’s “30th National College Information Application Service Innovation Competition” in 2025. These achievements demonstrate the team’s ability to translate AI technologies into practical applications that address pressing cybersecurity challenges, particularly in deepfake detection.

Reflecting on their development process, the student team highlighted the challenge of balancing detection performance with the computational constraints of mobile devices. From data processing and model tuning to system integration, each stage required close collaboration and iterative problem-solving. Through this hands-on experience, the students deepened their understanding of deep learning and multimodal technologies, while also gaining insight into the responsible application of AI. Looking ahead, the team plans to expand data sources and application scenarios, aiming to make their deepfake detection system a widely accessible tool for identifying manipulated media.

#SDG4 #SDG9 #SDG16