| Summary: | The advancement of computing power has facilitated the development of three-dimensional point cloud neural networks, and the issue of adversarial samples in point clouds has garnered increasing attention. However, Current adversarial training methods often result in reduced classification effectiveness of the model. Additionally, the scarcity of point cloud data poses challenges in acquiring sufficient training samples. This paper proposes a novel algorithm for classifying 3D point cloud adversarial samples, employing a combination of self-supervision and information gain. The algorithm first utilizes a self-supervised learning model to extract potential information from the data. Then, it employs information gain to screen the features of the adversarial samples and concentrate on the beneficial classification information. The selected features are fed back to the classification model through joint training, aiming to further enhance classification accuracy. In comparison to the standard training approaches of Pointnet, Pointnet++, and Dgcnn on the Modelnet40 and ScanObjectNN datasets, the proposed algorithm exhibits significant improvements in classification effectiveness. Specifically, on the ModelNet40 dataset, the algorithm achieves improvements of 0.70, 0.88, and 0.07 in classification effectiveness over PointNet, PointNet++, and DGCNN, respectively. Similarly, on the ScanObjectNN dataset, the algorithm demonstrates improvements of 0.06, 0.54, and 1.72 in classification effectiveness for the three networks, respectively. These results demonstrate that the algorithm is capable of overcoming the limitations of traditional adversarial training, which tends to reduce the generalization performance of the model, and further enhances the classification accuracy.
|
|---|