Fast Adversarial Training for Graph Neural Network Based Resource Allocation

In this letter, we propose a meta-learning-based fast adversarial training method to address the vulnerability of graph neural network (GNN) based resource allocation method to adversarial attacks. For a specific wireless network configuration, we design a convolutional neural network (CNN) based adversarial generator and perform adversarial training with the generated samples. Besides, we utilize a context-based meta-learning method to improve generalization of the adversarial training method across different configurations. Simulation results demonstrate that the proposed method can effectively accelerate the adversarial training process and reduce the required training samples, which is promising for enhancing the robustness of learning-based resource allocation schemes.