Optimizing Latent Variables in Integrating Transfer and Query Based Attack Framework.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning Security

Background:

• Black-box adversarial attacks face limitations: transfer-based attacks suffer from poor performance due to model architecture mismatches, while query-based attacks demand extensive queries and high-dimensional optimization.
• Existing methods struggle to balance attack efficacy with computational efficiency and adaptability in adversarial machine learning.

Purpose of the Study:

• To propose a novel, integrated framework for black-box adversarial attacks that combines the strengths of transfer-based and query-based approaches.
• To address the limitations of poor transferability and high query requirements in current adversarial attack methodologies.

Main Methods:

• A two-phase framework: first, training an adversarial generator using an adversarial loss function to output adversarial perturbations.
• Latent variables are introduced as generator inputs to reduce optimization dimensionality.
• Second phase utilizes particle swarm optimization on latent variables to refine perturbations for successful attacks, demonstrated on ImageNet.

Main Results:

• The proposed framework significantly outperforms several state-of-the-art black-box adversarial attack methods in terms of attack performance.
• Experiments on the ImageNet dataset validate the framework's effectiveness.
• The framework demonstrates flexibility through successful extension to few-pixel attack scenarios.

Conclusions:

• The novel integrated framework effectively overcomes the limitations of traditional transfer-based and query-based black-box attacks.
• The method offers improved attack performance and efficiency, with potential for broader applications in adversarial machine learning and robustness testing.