AI-UNet: Attention Information-based deep URL Network for adult webpage classification

Internet provides us a plethora of knowledge and information across different domains. However, the internet also consists of mature contents like pornographic websites that has to be filtered to keep the internet kids friendly. There is a demand for automated adult webpage classifier which can handle the ever-growing World Wide Web efficiently and swiftly. In this research work, we propose a deep neural network architecture called as Attention Information-based URL Network (AI-UNet) for classifying the adult web sites without fetching the contents of the web page and can aid in quick decision making. The proposed AI-UNet is a two-stage classifier that uses special attention mechanisms to classify the given webpage as “Adult” or “Non-Adult” by leveraging the information present in an URL along with its web page title with a novel URL representation. A novel dense vector representation called as word-as-character embedding is proposed in this work. This embedding combines the advantages of both character and word level embedding with an effective pre-processing technique. To extract the highly relevant information present either in URL/title/both, an adaptive attention mechanism is added on top of the proposed recurrent convolutional neural network model. Our experimental evaluations show that the proposed AI-UNet model outperforms all other character and word embedding-based webpage classifiers including the recent URLNet model. Our proposed model achieves 89.92% accuracy over the benchmark ODP dataset. Furthermore, we evaluated the model with our own independent dataset called as real-time dataset and obtained an accuracy of 99.43% on this corpus. From the extensive experimental analysis, it is shown that there is a significant improvement over the existing methods for Adult Webpage classification.