Improving the adversarial transferability with relational graphs ensemble adversarial attack.

Area of Science:

• Artificial Intelligence
• Machine Learning Security

Background:

• Transferable black-box attacks are crucial for evaluating unknown model vulnerabilities.
• Existing methods often overlook complex dependencies between models, leading to suboptimal attacks.

Purpose of the Study:

• To propose a novel approach, Relational Graph Ensemble Attack (RGEA), for exploiting inter-model dependencies.
• To enhance the transferability of adversarial attacks across multiple models.

Main Methods:

• Redefined multi-model ensemble attack as a multi-objective optimization problem.
• Developed a simplified sub-optimization problem using model vector representations and a dependency matrix.
• Theoretically analyzed the connection between RGEA and Multiple Gradient Descent Algorithm (MGDA).

Main Results:

• RGEA effectively exploits dependencies among multiple models, addressing unbalanced and inadequate attacks.
• The approach simplifies computationally intensive sub-optimization problems.
• Experiments on the LFW dataset showed improved white-box attack success rates.

Conclusions:

• RGEA significantly boosts the transferability of black-box adversarial attacks.
• The method enhances existing gradient-based attacks when combined.
• RGEA offers a more effective strategy for attacking ensembles of models.