Although current face recognition systems perform well in relatively constrained scenes, they are often affected by secondary creation of netizens, serious image blurring and abundant posture changes in real-world Internet videos. Focusing on these problems, in this paper, we propose a face recognition model names Internet Video-based Face Recognition Network (IVFRNet) based on deep learning for real Internet videos. And we propose a weighted loss function to enhance the ability of learned features. To test the model, we construct a small-scale real-world Internet video-based face dataset. The experiment results show that our method outperforms other methods.