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

The transferability of adversarial examples is crucial for achieving black-box attacks in real-world scenarios. However, existing methods often exhibit limited effectiveness in terms of attack success rates and transferability across different model architectures (e.g., from CNN-based architectures to Vision Transformer (ViT)-based architectures). Such limitations impede reliable adversarial evaluation in safety-critical applications, such as autonomous driving and medical imaging. To address these challenges, we propose FUGEA, which decouples the adversarial attack process into a two-stage optimization designed to enhance cross-architecture transferability and attack success rates. In the first stage, the Rapid Convergence Engine (RCE) integrates Uncertainty Weight (UW) and Gradient Agreement Mapping (GAM) to dynamically allocate ensemble gradient weights and stabilize the optimization direction. In the second stage, the Precision Refinement Gradient (PRG) module employs Sampled Neighbor Predictive Gradient (SNPG) to achieve fine-grained perturbation optimization through the fusion of neighborhood sampling and future-step prediction. Extensive experiments demonstrate that FUGEA consistently outperforms state-of-the-art methods in attack success rates and transferability across both CNN and ViT architectures. Furthermore, it maintains robust performance against advanced adversarial defenses, providing a powerful and effective solution for real-world cross-architecture black-box attacks.