Mejoras en ocultamiento de pertenencia a comunidades mediante una arquitectura especializada de aprendizaje por refuerzo

As complex networks face increasingly advanced algorithmic analysis, the ability to conceal specific structural affiliations has emerged as a relevant graph algorithmic problem. Current online platforms rely on sophisticated techniques to characterize and cluster users in order to provide more contextualized content and services. One such approach is based on community detection algorithms applied to large underlying graphs. In this context, users are characterized by preferences or attributes, which can reveal information that may be considered sensitive. Community Membership Hiding is the problem of strategically rewiring network topology to camouflage a target node and evade detection by clustering algorithms. While recent Deep Reinforcement Learning (DRL) approaches have pioneered this line of work, they often face efficacy and scalability issues on large-scale graphs. By dissecting state-of-the-art models, we identified a critical architectural limitation: agents evaluate the network globally without explicitly conditioning on the target node, while relying on static embeddings that fail to capture real-time topological changes during the rewiring process. In this paper, we introduce a novel DRL architecture designed to overcome these fundamental limitations. We propose an Actor-Critic system powered by Graph Attention Networks v2 (GATv2). By leveraging a target-aware feature set that combines static original-graph indicators with dynamically updated structural signals, and by explicitly contextualizing the target node through a dynamic attention mechanism, our model accurately represents the evolving graph state. Extensive experimental evaluation confirms that our architecture significantly outperforms existing baselines, achieving high hiding success rates with minimal structural perturbation and notable gains in computational efficiency.