Data Leakage Prediction Based on LSTM and Graph Convolutional Network: Dynamic Analysis of Sensitive Data Access Patterns

Abstract

In the era of information technology, data leakage poses a severe threat to the interests of individuals, businesses, and nations. In this study, we focus on predicting data leakage by combining LSTM (a specialized recurrent neural network that captures the time-series characteristics of sensitive data access behavior through mechanisms such as forget gates, input gates, and output gates) with graph-structured data and Graph Convolutional Networks (GCNs). GCNs excel at constructing relevant graphs to reveal the intricate relationships between data. The fusion of these two methods integrates time-series and spatial structural features, thereby enhancing prediction accuracy and timeliness. Through simulation experiments and practical applications in various fields such as healthcare and finance, we have demonstrated that this combined approach outperforms single models. To address the challenges of high model complexity, data complexity and diversity, as well as data security and privacy, we have employed optimization algorithms, advanced feature extraction techniques, and methods such as differential privacy and homomorphic encryption. This research provides robust support for data security protection and will continue to explore and refine related technologies in the future.