arXiv:2606.08067v1 Announce Type: new 
Abstract: Graph neural networks (GNNs) are widely deployed on relational data, yet they can leak sensitive or proprietary information about the training graph adjacency, e.g., social ties, transactions, and interactions. This work studies graph reconstruction attacks (GRA), a form of model inversion that reconstructs the training adjacency from a trained GNN, given different levels of attacker-side information. We first provide a systematic characterization of when and why adjacency becomes recoverable through features, labels, embeddings, and predictions, with leakage modulated by graph homophily, heterophily, and the model's inductive bias. Motivated by these findings, we view GNN inference through a Markov chain approximation lens, treating the layered forward computation as a chain of topology-dependent representations. Building on this view, we develop complementary attack and defense methods. On the attack side, we propose MC-GRA (+), which reconstructs the adjacency by optimizing a surrogate adjacency whose GNN-induced representations align with those of the target model at each layer. On the defense side, we propose MC-GPB (+), which suppresses adjacency-dependent information throughout the representation chain while aiming to preserve classification accuracy under a privacy-utility trade-off. Experiments across homophilic/heterophilic graph benchmarks and GNNs show that our attacks improve reconstruction fidelity over prior methods, while our defenses reduce reconstruction success with only minor accuracy loss.

أظهرت الأبحاث الحديثة نقاط الضعف في الشبكات العصبية البيانية (GNN) أمام هجمات إعادة بناء الرسوم البيانية، والتي يمكن أن تكشف عن بيانات تدريب حساسة. تصف هذه الدراسة بشكل منهجي الظروف التي يمكن من خلالها استعادة معلومات الاتصال، المتأثرة بعوامل مثل التماثل في الرسوم البيانية والانحياز الاستقرائي للنموذج. كما تقدم MC-GRA، وهي طريقة هجوم جديدة، إلى جانب استراتيجيات دفاعية مكملة لتخفيف هذه المخاطر.

Investigaciones recientes han destacado las vulnerabilidades de las Redes Neuronales Gráficas (GNN) ante ataques de reconstrucción de grafos, que pueden exponer datos sensibles de entrenamiento. Este estudio caracteriza sistemáticamente las condiciones bajo las cuales se puede recuperar información de adyacencia, influenciada por factores como la homofilia del grafo y el sesgo inductivo del modelo. Introduce MC-GRA, un nuevo método de ataque, junto con estrategias de defensa complementarias para mitigar estos riesgos.

Des recherches récentes ont mis en évidence les vulnérabilités des réseaux de neurones graphiques (GNN) face aux attaques de reconstruction de graphes, qui peuvent exposer des données sensibles d'entraînement. Cette étude caractérise systématiquement les conditions dans lesquelles les informations d'adjacence peuvent être récupérées, influencées par des facteurs tels que l'homophilie des graphes et le biais inductif du modèle. Elle introduit MC-GRA, une nouvelle méthode d'attaque, ainsi que des stratégies de défense complémentaires pour atténuer ces risques.

Recent research has highlighted the vulnerabilities of Graph Neural Networks (GNNs) to graph reconstruction attacks, which can expose sensitive training data. This study systematically characterizes the conditions under which adjacency information can be recovered, influenced by factors such as graph homophily and the model's inductive bias. It introduces MC-GRA, a novel attack method, alongside complementary defense strategies to mitigate these risks.