Research Direction: Media Security / Trustworthy Media Forensics
Authors: C.-Y. Lai, C.-Y. Jian, P.-C. Chuang, C.-M. Lee, C.-C. Hsu, C.-T. Hsu, and C.-W. Lin
Our paper UMCL: Unimodal-Generated Multimodal Contrastive Learning for Cross-compression-rate Deepfake Detection has been published in the International Journal of Computer Vision (IJCV).
UMCL studies a practical but difficult DeepFake detection setting: models often degrade when the compression level of the test video differs from that of the training data. To address this cross-compression-rate challenge, the paper derives multiple complementary cues from a single visual input, including rPPG signals, facial landmark dynamics, and semantic embeddings, and aligns them through contrastive learning for more robust detection.
This work reflects our lab’s ongoing efforts in trustworthy media analysis, robust vision under real-world degradations, and security-oriented visual intelligence that remains reliable beyond curated benchmark conditions.
