MUGEN Protects Graph Data from Unauthorized AI Learning Across Tasks

Researchers have developed MUGEN, a novel framework designed to safeguard sensitive graph data from unauthorized machine learning applications. Unlike previous methods that protect against a single specific task, MUGEN creates a single, feature-perturbed dataset that renders the data "unlearnable" for a variety of common graph learning tasks simultaneously, including node classification, graph classification, and link prediction. This comprehensive protection is achieved through a shared Graph Neural Network (GNN) encoder and task-specific heads, ensuring that models trained on the perturbed data fail to generalize to clean data. The framework employs a Task-Aligned Separability Objective (TASO) to enhance unlearnability and its transferability across different GNN architectures and tasks. Additionally, Type-Adaptive Perturbation (TAP) customizes the perturbation process based on node attribute types, allowing for effective protection of both discrete and continuous node features. Experimental results across multiple benchmarks and GNN backbones demonstrate MUGEN's effectiveness in generating transferable unlearnable graph examples, even under adversarial training conditions.