Deep learning-based multi-tier sensitivity analysis network for document sensitivity classification

Abstract

In the digital age, the exponential growth of data necessitates robust and efficient systems for document classification to maintain data security and compliance. Text classification plays a crucial role in identifying sensitive information by automatically categorizing documents based on their content. Using advanced machine learning and deep learning models, it analyzes text to detect keywords, patterns, and contextual cues that indicate the presence of sensitive data. This paper presents a novel framework, the multi-tier sensitivity analysis network (MTSAN), designed to accurately classify documents into public, private, and confidential categories. The proposed system integrates several advanced components, including the multi-tier sensitivity encoding network (MTSEN). MTSAN leverages a combination of convolutional networks and graph convolutional networks (GCNs) to capture both local and global contextual information. The dual-scope graph convolution block (DSGCB) is introduced to address both global dependencies and local dynamics, employing a novel fusion mechanism to merge global and local features effectively. Additionally, the cross-tier information fusion block (CTIFB) facilitates the seamless integration of multi-level features, further refining the classification process. The results demonstrate that the proposed MTSAN model outperforms traditional machine learning approaches and contemporary deep learning models such as bidirectional encoder representations from transformers (BERT), achieving superior accuracy and F1 scores in classifying sensitive information.