Automatic Security Classification Based on Incremental Learning and Similarity Comparison

摘要

Document security classification is the foundation of security management for the sensitive and confidential information. Different with the general document classification, the security one is more difficult and challenging, due to the diverse and changeable sensitive features and high requirement for accuracy. In this paper, we propose a new method named incremental learning and similarity comparison (ILSC), which combines two effective security labeling strategies for automatic security classification. In fact, the sensitive document dataset augments continuously. Accordingly, we exploit the use of incremental learning to capture continuously the useful information of the new classified documents and update the classifier. In addition, to utilize the identified sensitive sentences in the classified documents, we introduce a method based on similarity comparison of sentence features, as a supplement to the prediction obtained by the incremental learning. Experimental results show the proposed method can produce competitive results in terms of accuracy.