Derivative-free optimization adversarial attacks for graph convolutional networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Graph convolutional networks (GCNs) excel at graph data processing but are vulnerable to adversarial attacks.
• Adversarial attacks can manipulate graph structures (edges or nodes) to compromise GCN classification performance.

Purpose of the Study:

• To propose a novel black-box adversarial attack framework for GCNs using derivative-free optimization (DFO).
• To implement a direct attack algorithm (DFDA) and optimize its search space for effective graph adversarial example generation.

Main Methods:

• Developed a black-box adversarial attack framework leveraging derivative-free optimization (DFO) algorithms.
• Implemented the Direct Function Derivative Attack (DFDA) using the Nevergrad library.
• Redesigned the perturbation vector using constraint size to manage the large search space.

Main Results:

• The DFDA framework successfully generated graph adversarial examples without relying on gradient information.
• DFDA demonstrated superior performance compared to Nettack in most experimental scenarios.
• Achieved over 95% average attack success rate on the Cora dataset with minimal edge perturbations (at most eight).

Conclusions:

• The proposed DFO-based framework effectively exploits derivative-free optimization for robust node classification adversarial attacks.
• DFDA offers a promising approach to understanding and mitigating vulnerabilities in GCNs against adversarial manipulation.