CMU LSMM Project: Deepfake Detection
This project is a part of Large-Scale Multimedia Analysis (11-775, Spring 2020) class at CMU. Objective of our team is to improve the detection model specialized on fake facial reenactment images generated by GAN-based method (NeuralTexture)
This project is part of High-Potential Individuals Global Training Program supported by IITP Korea. The purpose of this project is to learn and implement hands-on data analysis, AI / ML, and software engineering in real-world situations. Our team deepfake dectector through this project.
- Facial reenactment is manipulation of human facial expression using some 'actor' source, which can be serious problem if it is abused upon renowned individuals (like the president of the United Stated, etc.)
- From FaceForensics++ paper, we found that however, it is very hard to detect fake facial reenactment image from GAN for automated model, and even for human perspective. In addition, we are going to try improve detection performance upon low-quality fake images.
- Therefore, we focus on the facial reenactment detection for social goodness, based on the state-of-the-art CNN-based image classifier, EfficientNet. Also, we implement the recent semi-supervised learning approach with noise added model, so called the "Noisy Student Training“.
Programming Language & Technology Specification
This research was results of a study on the “HPC (High Performance Computing) Support” Project, supported by the ‘Ministry of Science and ICT (Information and Communication Technology)’ and NIPA (National IT Industry Promotion Agency), Repulbic of Korea.