Generating Adversarial Examples for 3D Point Clouds Using Evolutionary Algorithms

We propose a general attack framework based on evolutionary algorithms to quickly and efficiently generate low-perturbation adversarial samples for 3D point cloud data. Specifically, we employ a random selection to increase or decrease the number of points perturbed, followed by a single-objective evolutionary algorithm to adjust the magnitude of the perturbations. Additionally, we design an extended version that introduces a multi-objective optimization algorithm to simultaneously optimize the magnitude of the perturbations and attack effectiveness. To evaluate the performance of the proposal, we compare it with other classic algorithms using the ModelNet40 dataset. Experimental results show that the attack algorithm based on evolutionary algorithms can successfully fool well-trained models with acceptable perturbations and a higher success rate.