FUGEA: Fused unified gradient ensemble for cross-architecture transferable attacks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Adversarial attacks are vital for evaluating AI robustness, especially in safety-critical fields.
• Current black-box attack methods struggle with cross-architecture transferability (e.g., CNN to ViT) and low success rates.
• Limitations hinder reliable adversarial evaluation in autonomous driving and medical imaging.

Purpose of the Study:

• To develop a novel adversarial attack method, FUGEA, enhancing cross-architecture transferability and attack success rates.
• To address the limitations of existing methods in black-box attacks across diverse model architectures.
• To provide a robust solution for real-world adversarial evaluation.

Main Methods:

• FUGEA employs a two-stage optimization process: Rapid Convergence Engine (RCE) and Precision Refinement Gradient (PRG).
• RCE integrates Uncertainty Weight (UW) and Gradient Agreement Mapping (GAM) for stabilized optimization and dynamic gradient weighting.
• PRG utilizes Sampled Neighbor Predictive Gradient (SNPG) for fine-grained perturbation optimization via neighborhood sampling and prediction.

Main Results:

• FUGEA significantly outperforms state-of-the-art methods in attack success rates and transferability across CNN and ViT architectures.
• The proposed method demonstrates robust performance against advanced adversarial defenses.
• FUGEA provides effective cross-architecture black-box attacks for real-world applications.

Conclusions:

• FUGEA offers a powerful and effective solution for cross-architecture black-box adversarial attacks.
• The two-stage optimization strategy successfully enhances transferability and attack efficacy.
• This research contributes to more reliable adversarial evaluation in safety-critical AI systems.