Development of novel approaches based on geometric deep learning for graph visualization of cyber attacks

The cyber threats to which the digital ecosystem is exposed are exhibiting increasing complexity and intricacy. These attacks are not isolated incidents, but multilayered and systematic structures. This structure creates a foundation where traditional Euclidean data models fall short. This thesis presents a new approach by modeling cyber threat intelligence with graph structures. The aim is to move beyond Euclidean deep learning models and provide a new perspective on the early detection of cyberattacks and the development of proactive defense strategies. To this end, the thesis presents two fundamental approaches for detecting and visualizing cyberattacks. The first contribution is the development of an interactive visualization and analysis architecture based on the Graph Attention Network for the detection and analysis of cyberattacks. This system transforms raw data from a static body of information into a dynamic and intuitive discovery environment for security analysts. Through this platform, analysts can monitor the temporal evolution of attacks, compare threat vectors, and develop defensive strategies based on concrete data. At the heart of this process is the transformation of data obtained through meticulous preprocessing and feature engineering into a graph structure reflecting meaningful relationships. The resulting structure is presented in a dynamic interface using the JavaScript-based SigmaJS library, allowing analysts to directly interact with the data. This integrates advanced machine learning techniques with a human-centered visualization experience, establishing a new standard for interpreting cyber threat intelligence. The second contribution is the presentation of a hybrid model focused on predicting cyberattacks before they occur. For this purpose, a three-layer Dynamic Graph Neural Network (DGNN) architecture is proposed, combining the Gated Recurrent Unit (GRU) structure with a meta-path-based attention mechanism. The uniqueness of the model lies in its inclusion of attention-based embedding layers at both the node and meta-path levels and a dedicated module that learns evolutionary patterns. This enables the model to deeply analyze structural and temporal correlations in complex and constantly changing graph data. Initial results show that the model predicts attacks with high accuracy and significantly reduces the false alarm rate. The overall conclusion of the thesis is that graph-based AI significantly enhances predictive capabilities and intuitive data visualization in cybersecurity.