Adversarial attacks are an effective method for revealing the vulnerabilities of 3D point cloud classification models and promoting the development of more robust architectures. However, existing gradient-based attack methods often exhibit unstable update directions and slow convergence, primarily arising from the complex loss landscape associated with the unordered, sparse, and irregular structure of point cloud data, and this instability in turn compromises the imperceptibility of the generated adversarial examples. To address these challenges, a novel two-stage gradient optimization framework is proposed to generate adversarial point clouds with improved imperceptibility and optimization efficiency. Specifically, we introduce a Hybrid Gradient Descent (HGD) strategy that applies geometric transformations to create augmented samples around each adversarial point cloud. By aggregating the gradients from augmented samples, HGD effectively smooths local noise and guides the optimization toward more stable descent directions. Extending the HGD strategy, we design a Prospective Gradient Correction (PGC) mechanism that constructs forward-looking perturbation points along the current update trajectory and fuses the corresponding gradients with the original direction. PGC enables dynamic refinement of the update path, mitigating local gradient bias and aligning the update direction more accurately with the decision boundary. Experimental results show that the full HGD&PGC framework reduces required iterations by 48.5% and achieves a reduction of 18.7% in the average perturbation magnitude, all without compromising attack success rates. Furthermore, this optimization framework has the potential to be extended to other 3D data types, and also to inform defense strategies that mitigate such attacks by complicating the loss landscape.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号