This paper presents a large-scale face dataset, CASIA-WebFace, collected from the Internet, containing 10,575 subjects and 494,414 face images. The dataset was created to address the lack of large-scale public face datasets, which is critical for evaluating and improving face recognition algorithms. The dataset was constructed using a semi-automatic method, leveraging the structured information on the IMDb website to crawl and annotate face images. The dataset is used to train a deep convolutional neural network (CNN) to learn discriminative face representations, achieving state-of-the-art performance on the LFW and YTF datasets. The proposed CNN architecture includes 10 convolutional layers, 5 pooling layers, and 1 fully connected layer, with a combination of recent techniques such as ReLU neurons, dropout, and multiple loss functions. The network is evaluated on LFW and YTF, achieving performance comparable to existing state-of-the-art methods like DeepFace. The paper also discusses the importance of large-scale datasets in face recognition research and highlights the potential of CASIA-WebFace to advance the field by providing a standardized benchmark for evaluating face recognition algorithms in unconstrained environments. The dataset and the proposed CNN are made publicly available to encourage further research and development in face recognition technology.This paper presents a large-scale face dataset, CASIA-WebFace, collected from the Internet, containing 10,575 subjects and 494,414 face images. The dataset was created to address the lack of large-scale public face datasets, which is critical for evaluating and improving face recognition algorithms. The dataset was constructed using a semi-automatic method, leveraging the structured information on the IMDb website to crawl and annotate face images. The dataset is used to train a deep convolutional neural network (CNN) to learn discriminative face representations, achieving state-of-the-art performance on the LFW and YTF datasets. The proposed CNN architecture includes 10 convolutional layers, 5 pooling layers, and 1 fully connected layer, with a combination of recent techniques such as ReLU neurons, dropout, and multiple loss functions. The network is evaluated on LFW and YTF, achieving performance comparable to existing state-of-the-art methods like DeepFace. The paper also discusses the importance of large-scale datasets in face recognition research and highlights the potential of CASIA-WebFace to advance the field by providing a standardized benchmark for evaluating face recognition algorithms in unconstrained environments. The dataset and the proposed CNN are made publicly available to encourage further research and development in face recognition technology.