Dual-Task Cascaded for Proactive Deepfake Detection Using QPCET Watermarking

In recent years, the rapid development of deepfake technology has brought serious threats to facial information security. To address this issue, numerous passive deepfake detection methods have been developed with notable success. However, these methods are limited in offering proactive defense against deepfakes. To this end, we propose a proactive deepfake detection framework that integrates Quaternion Polar Complex Exponential Transform (QPCET) with deep learning, treating watermark embedding and extraction as two separate processes. Firstly, we embed the watermark information into QPCET coefficients, enhancing the robustness of the watermark against conventional attacks while ensuring its imperceptibility. Secondly, we propose a Dual-Task Cascaded Detection (DTCD) framework for watermark extraction and deepfake detection. Additionally, we introduce a Self-Attention Moment-Aware Watermark Detection (SAM-WD) module, which aids the model in more accurately perceiving watermark embedding regions. Through knowledge distillation between the two tasks, the model can accurately extract watermarks under conventional attacks and accurately detect deepfake. Experimental results on benchmark datasets demonstrate that our method surpasses state-of-the-art techniques, delivering exceptional performance in watermark robustness and imperceptibility while simultaneously accomplishing accurate deepfake detection.