Deepfake Detection for Online Media

With the rise of AI-generated content, deepfakes pose serious threats to media integrity, politics, and cybersecurity. Vietnam, like many countries, faces growing risks of misinformation and fraud through manipulated images. This project specifically focuses on detecting deepfakes in Vietnamese facial images to address the lack of localized solutions.



The project introduces a deepfake detection system built with Convolutional Neural Networks (CNN) enhanced by Squeeze-and-Excitation (SE) blocks. The model is trained on a large, diverse dataset that combines international benchmarks (FaceForensics++, Celeb-DF, DFDC) with a Vietnamese dataset generated using StyleGAN, allowing it to capture diverse patterns while still maintaining focus on Vietnamese data.

The system achieves over 90% detection accuracy with strong precision-recall performance and low false positive rates, especially for the Vietnamese test dataset. Deployed as a web platform and browser extension, it supports real-time inference where users can upload images and instantly verify authenticity. The result is a fast and user-friendly tool that demonstrates robust detection across multiple deepfake generation techniques.