<p>With the rapid advancement of artificial intelligence, DeepFake has provided novel visual effects and artistic creation methods across various fields. However, highly realistic face forgery videos pose severe threats to individual privacy, social stability, and national security. To address these challenges, effective forgery detection methods are crucial for mitigating facial security risks. The current research focuses on detecting authenticity by personalized features, which are based on the distribution regularity of real and fake videos under a particular dataset. But these features rely heavily on the pixel domain of the original videos and the acquisition of a priori information. Therefore, in this study, we turn to non-personalized features, which are based on the universal regularity that is available in reality and can be applied to different videos and scenarios. Then, we establish a non-personalized feature extraction framework based on remote photoplethysmograph (rPPG) physiological signal to detect forged videos and also propose an assessment method for non-personalized feature of physiological signal. Furthermore, extensive experiments have validated the feasibility of our proposed method. Compared to existing methods, our approach demonstrates good generalization capabilities on various DeepFake datasets.</p>