| Abstract: |
| In real-world black-box attack scenarios, transferability is the primary threat vector. Targeted attacks, while of high practical importance, are notoriously difficult to transfer across models with heterogeneous architectures. This work introduces a Probability Manifold Framework that quantifies the intrinsic density of high-dimensional data by implicitly modeling the geometry of its underlying low-dimensional manifold. We theoretically prove that our density measure monotonically reflects the distance from a sample to the data manifold and demonstrate that guiding adversarial perturbations toward High-Sample-Density Regions (HSDR) is the optimal strategy for improving targeted transferability. To generalize this guidance to arbitrary feature topologies, we devise a topology-aware Multi-Anchor Softmin Strategy. This approach enables adaptive matching to valid high-density modes instead of forcing a rigid approximation to a single centroid, thereby avoiding low-density voids and accelerating convergence. Building on these innovations, we propose MAGMA (Multi-Anchor Generative Manifold Attack), a novel generative method for targeted attacks. Extensive experiments on ImageNet show that MAGMA sets a new state-of-the-art for generative attacks, substantially outperforming existing methods (e.g., TTP, ESMA) in average transfer success rate while also achieving significant gains in training efficiency. |
|