Robustness meets accuracy in adversarial training for graph autoencoder

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph Autoencoders (GAE) are effective for graph embedding but susceptible to adversarial attacks.
• Adversarial training enhances GAE robustness but can degrade natural accuracy.
• Balancing robustness and natural accuracy is critical for GAE performance.

Purpose of the Study:

• To propose an improved GAE model combining Structure and Feature encoders.
• To introduce a novel adversarial training strategy (GAE-SFAT) for enhanced GAE robustness and accuracy.
• To develop an optimization algorithm for GAE-SFAT considering both robustness and accuracy.

Main Methods:

• Formulated an improved GAE by integrating Structure and Feature encoders.
• Developed GAE-SFAT with a refined adversarial scope for training.
• Designed a novel optimization algorithm tailored for GAE-SFAT.

Main Results:

• GAE-SFAT demonstrated improved robustness against adversarial attacks.
• The proposed method mitigated the degradation of natural accuracy compared to standard adversarial training.
• Experiments on three datasets showed GAE-SFAT outperformed state-of-the-art adversarial training models under various perturbations.

Conclusions:

• GAE-SFAT offers a superior approach to adversarial training for graph autoencoders.
• The method effectively balances model robustness and natural accuracy.
• GAE-SFAT represents a significant advancement in secure and accurate graph representation learning.